diff --git "a/judgment.mlmodelc/model.mil" "b/judgment.mlmodelc/model.mil"
--- "a/judgment.mlmodelc/model.mil"
+++ "b/judgment.mlmodelc/model.mil"
@@ -2217,2209 +2217,2209 @@ program(1.3)
             string var_5445_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5445_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
             tensor<fp32, [1, 16]> logits = cast(dtype = var_5445_cast_fp16_to_fp32_dtype_0, x = var_5445_cast_fp16)[name = string("cast_489")];
         } -> (logits);
-    func len256<ios18>(tensor<int32, [1, 256]> attention_mask, tensor<int32, [1, 256]> input_ids, tensor<int32, [1, 16]> label_positions) {
-            tensor<int32, [1, 1, 256]> slots = const()[name = string("slots"), val = tensor<int32, [1, 1, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922400896)))];
+    func len192<ios18>(tensor<int32, [1, 192]> attention_mask, tensor<int32, [1, 192]> input_ids, tensor<int32, [1, 16]> label_positions) {
+            tensor<int32, [1, 1, 192]> slots = const()[name = string("slots"), val = tensor<int32, [1, 1, 192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922400896)))];
             int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
             bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
             tensor<fp16, [128011, 1024]> encoder_embeddings_word_embeddings_weight_to_fp16 = const()[name = string("encoder_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [128011, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640)))];
             int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
-            tensor<bool, [1, 256]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            tensor<bool, [1, 192]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
             int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128011)];
-            tensor<int32, [1, 256]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
-            tensor<int32, [1, 256]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            tensor<int32, [1, 192]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1, 192]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
             int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
-            tensor<fp16, [1, 256, 1024]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0, validate_indices = inputs_embeds_validate_indices_0, x = encoder_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0, validate_indices = inputs_embeds_validate_indices_0, x = encoder_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")];
             tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_embeddings_LayerNorm_weight_to_fp16 = const()[name = string("encoder_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262167232)))];
             tensor<fp16, [1024]> encoder_embeddings_LayerNorm_bias_to_fp16 = const()[name = string("encoder_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262169344)))];
             fp16 var_31_to_fp16 = const()[name = string("op_31_to_fp16"), val = fp16(0x1p-23)];
-            tensor<fp16, [1, 256, 1024]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = encoder_embeddings_LayerNorm_bias_to_fp16, epsilon = var_31_to_fp16, gamma = encoder_embeddings_LayerNorm_weight_to_fp16, x = inputs_embeds_cast_fp16)[name = string("embeddings_1_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = encoder_embeddings_LayerNorm_bias_to_fp16, epsilon = var_31_to_fp16, gamma = encoder_embeddings_LayerNorm_weight_to_fp16, x = inputs_embeds_cast_fp16)[name = string("embeddings_1_cast_fp16")];
             tensor<int32, [1]> mask_3_axes_0 = const()[name = string("mask_3_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<int32, [1, 256, 1]> mask_3 = expand_dims(axes = mask_3_axes_0, x = attention_mask)[name = string("mask_3")];
+            tensor<int32, [1, 192, 1]> mask_3 = expand_dims(axes = mask_3_axes_0, x = attention_mask)[name = string("mask_3")];
             string mask_to_fp16_dtype_0 = const()[name = string("mask_to_fp16_dtype_0"), val = string("fp16")];
-            tensor<fp16, [1, 256, 1]> mask_3_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = mask_3)[name = string("cast_492")];
-            tensor<fp16, [1, 256, 1024]> input_1_cast_fp16 = mul(x = embeddings_1_cast_fp16, y = mask_3_to_fp16)[name = string("input_1_cast_fp16")];
-            tensor<int32, [3]> var_49 = const()[name = string("op_49"), val = tensor<int32, [3]>([1, 1, 256])];
-            tensor<int32, [1, 1, 256]> var_50 = reshape(shape = var_49, x = attention_mask)[name = string("op_50")];
-            tensor<int32, [3]> attention_mask_3_reps_0 = const()[name = string("attention_mask_3_reps_0"), val = tensor<int32, [3]>([1, 256, 1])];
-            tensor<int32, [1, 256, 256]> attention_mask_3 = tile(reps = attention_mask_3_reps_0, x = var_50)[name = string("attention_mask_3")];
+            tensor<fp16, [1, 192, 1]> mask_3_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = mask_3)[name = string("cast_492")];
+            tensor<fp16, [1, 192, 1024]> input_1_cast_fp16 = mul(x = embeddings_1_cast_fp16, y = mask_3_to_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> var_49 = const()[name = string("op_49"), val = tensor<int32, [3]>([1, 1, 192])];
+            tensor<int32, [1, 1, 192]> var_50 = reshape(shape = var_49, x = attention_mask)[name = string("op_50")];
+            tensor<int32, [3]> attention_mask_3_reps_0 = const()[name = string("attention_mask_3_reps_0"), val = tensor<int32, [3]>([1, 192, 1])];
+            tensor<int32, [1, 192, 192]> attention_mask_3 = tile(reps = attention_mask_3_reps_0, x = var_50)[name = string("attention_mask_3")];
             int32 var_56 = const()[name = string("op_56"), val = int32(-1)];
             tensor<int32, [1]> attention_mask_5_axes_0 = const()[name = string("attention_mask_5_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [1, 1, 256, 256]> attention_mask_5 = expand_dims(axes = attention_mask_5_axes_0, x = attention_mask_3)[name = string("attention_mask_5")];
+            tensor<int32, [1, 1, 192, 192]> attention_mask_5 = expand_dims(axes = attention_mask_5_axes_0, x = attention_mask_3)[name = string("attention_mask_5")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262171456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264268672)))];
-            tensor<fp16, [1, 256, 1024]> linear_0_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
-            tensor<int32, [4]> var_189 = const()[name = string("op_189"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_3_cast_fp16 = reshape(shape = var_189, x = linear_0_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_0_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [4]> var_189 = const()[name = string("op_189"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_3_cast_fp16 = reshape(shape = var_189, x = linear_0_cast_fp16)[name = string("x_3_cast_fp16")];
             tensor<int32, [4]> var_191 = const()[name = string("op_191"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_192_cast_fp16 = transpose(perm = var_191, x = x_3_cast_fp16)[name = string("transpose_143")];
-            tensor<fp16, [16, 256, 64]> query_layer_3_cast_fp16 = reshape(shape = var_196, x = var_192_cast_fp16)[name = string("query_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_192_cast_fp16 = transpose(perm = var_191, x = x_3_cast_fp16)[name = string("transpose_143")];
+            tensor<fp16, [16, 192, 64]> query_layer_3_cast_fp16 = reshape(shape = var_196, x = var_192_cast_fp16)[name = string("query_layer_3_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264270784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266368000)))];
-            tensor<fp16, [1, 256, 1024]> linear_1_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
-            tensor<int32, [4]> var_203 = const()[name = string("op_203"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_7_cast_fp16 = reshape(shape = var_203, x = linear_1_cast_fp16)[name = string("x_7_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_1_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_203 = const()[name = string("op_203"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_7_cast_fp16 = reshape(shape = var_203, x = linear_1_cast_fp16)[name = string("x_7_cast_fp16")];
             tensor<int32, [4]> var_205 = const()[name = string("op_205"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_206_cast_fp16 = transpose(perm = var_205, x = x_7_cast_fp16)[name = string("transpose_142")];
-            tensor<fp16, [16, 256, 64]> key_layer_1_cast_fp16 = reshape(shape = var_210, x = var_206_cast_fp16)[name = string("key_layer_1_cast_fp16")];
+            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_206_cast_fp16 = transpose(perm = var_205, x = x_7_cast_fp16)[name = string("transpose_142")];
+            tensor<fp16, [16, 192, 64]> key_layer_1_cast_fp16 = reshape(shape = var_210, x = var_206_cast_fp16)[name = string("key_layer_1_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266370112)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268467328)))];
-            tensor<fp16, [1, 256, 1024]> linear_2_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
-            tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_11_cast_fp16 = reshape(shape = var_217, x = linear_2_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_2_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_11_cast_fp16 = reshape(shape = var_217, x = linear_2_cast_fp16)[name = string("x_11_cast_fp16")];
             tensor<int32, [4]> var_219 = const()[name = string("op_219"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_224 = const()[name = string("op_224"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_220_cast_fp16 = transpose(perm = var_219, x = x_11_cast_fp16)[name = string("transpose_141")];
-            tensor<fp16, [16, 256, 64]> value_layer_1_cast_fp16 = reshape(shape = var_224, x = var_220_cast_fp16)[name = string("value_layer_1_cast_fp16")];
+            tensor<int32, [3]> var_224 = const()[name = string("op_224"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_220_cast_fp16 = transpose(perm = var_219, x = x_11_cast_fp16)[name = string("transpose_141")];
+            tensor<fp16, [16, 192, 64]> value_layer_1_cast_fp16 = reshape(shape = var_224, x = var_220_cast_fp16)[name = string("value_layer_1_cast_fp16")];
             tensor<int32, [3]> var_234_perm_0 = const()[name = string("op_234_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_236_y_0_to_fp16 = const()[name = string("_inversed_236_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_234_cast_fp16 = transpose(perm = var_234_perm_0, x = key_layer_1_cast_fp16)[name = string("transpose_140")];
-            tensor<fp16, [16, 64, 256]> _inversed_236_cast_fp16 = mul(x = var_234_cast_fp16, y = _inversed_236_y_0_to_fp16)[name = string("_inversed_236_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_234_cast_fp16 = transpose(perm = var_234_perm_0, x = key_layer_1_cast_fp16)[name = string("transpose_140")];
+            tensor<fp16, [16, 64, 192]> _inversed_236_cast_fp16 = mul(x = var_234_cast_fp16, y = _inversed_236_y_0_to_fp16)[name = string("_inversed_236_cast_fp16")];
             bool attention_scores_1_transpose_x_0 = const()[name = string("attention_scores_1_transpose_x_0"), val = bool(false)];
             bool attention_scores_1_transpose_y_0 = const()[name = string("attention_scores_1_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = query_layer_3_cast_fp16, y = _inversed_236_cast_fp16)[name = string("attention_scores_1_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = query_layer_3_cast_fp16, y = _inversed_236_cast_fp16)[name = string("attention_scores_1_cast_fp16")];
             bool c2p_att_1_transpose_x_0 = const()[name = string("c2p_att_1_transpose_x_0"), val = bool(false)];
             bool c2p_att_1_transpose_y_0 = const()[name = string("c2p_att_1_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_292_to_fp16 = const()[name = string("op_292_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268469440)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_1_cast_fp16 = matmul(transpose_x = c2p_att_1_transpose_x_0, transpose_y = c2p_att_1_transpose_y_0, x = query_layer_3_cast_fp16, y = var_292_to_fp16)[name = string("c2p_att_1_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_1_cast_fp16 = matmul(transpose_x = c2p_att_1_transpose_x_0, transpose_y = c2p_att_1_transpose_y_0, x = query_layer_3_cast_fp16, y = var_292_to_fp16)[name = string("c2p_att_1_cast_fp16")];
             bool c2p_att_3_validate_indices_0 = const()[name = string("c2p_att_3_validate_indices_0"), val = bool(false)];
-            tensor<uint16, [16, 256, 256]> var_302_to_uint16 = const()[name = string("op_302_to_uint16"), val = tensor<uint16, [16, 256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922401984)))];
-            tensor<fp16, [16, 256, 256]> c2p_att_3_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_3_validate_indices_0, x = c2p_att_1_cast_fp16)[name = string("c2p_att_3_cast_fp16_cast_int16")];
+            tensor<uint16, [16, 192, 192]> var_302_to_uint16 = const()[name = string("op_302_to_uint16"), val = tensor<uint16, [16, 192, 192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922401728)))];
+            tensor<fp16, [16, 192, 192]> c2p_att_3_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_3_validate_indices_0, x = c2p_att_1_cast_fp16)[name = string("c2p_att_3_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_305_y_0_to_fp16 = const()[name = string("_inversed_305_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_305_cast_fp16 = mul(x = c2p_att_3_cast_fp16_cast_int16, y = _inversed_305_y_0_to_fp16)[name = string("_inversed_305_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_305_cast_fp16 = mul(x = c2p_att_3_cast_fp16_cast_int16, y = _inversed_305_y_0_to_fp16)[name = string("_inversed_305_cast_fp16")];
             bool p2c_att_1_transpose_x_0 = const()[name = string("p2c_att_1_transpose_x_0"), val = bool(false)];
             bool p2c_att_1_transpose_y_0 = const()[name = string("p2c_att_1_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_320_to_fp16 = const()[name = string("op_320_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270042432)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_1_cast_fp16 = matmul(transpose_x = p2c_att_1_transpose_x_0, transpose_y = p2c_att_1_transpose_y_0, x = key_layer_1_cast_fp16, y = var_320_to_fp16)[name = string("p2c_att_1_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_1_cast_fp16 = matmul(transpose_x = p2c_att_1_transpose_x_0, transpose_y = p2c_att_1_transpose_y_0, x = key_layer_1_cast_fp16, y = var_320_to_fp16)[name = string("p2c_att_1_cast_fp16")];
             bool var_328_validate_indices_0 = const()[name = string("op_328_validate_indices_0"), val = bool(false)];
-            tensor<uint16, [16, 256, 256]> var_327_to_uint16 = const()[name = string("op_327_to_uint16"), val = tensor<uint16, [16, 256, 256]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(924499200)))];
-            tensor<fp16, [16, 256, 256]> var_328_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_328_validate_indices_0, x = p2c_att_1_cast_fp16)[name = string("op_328_cast_fp16_cast_int16")];
+            tensor<uint16, [16, 192, 192]> var_327_to_uint16 = const()[name = string("op_327_to_uint16"), val = tensor<uint16, [16, 192, 192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(923581440)))];
+            tensor<fp16, [16, 192, 192]> var_328_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_328_validate_indices_0, x = p2c_att_1_cast_fp16)[name = string("op_328_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_3_perm_0 = const()[name = string("p2c_att_3_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_331_y_0_to_fp16 = const()[name = string("_inversed_331_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_3_cast_fp16 = transpose(perm = p2c_att_3_perm_0, x = var_328_cast_fp16_cast_int16)[name = string("transpose_139")];
-            tensor<fp16, [16, 256, 256]> _inversed_331_cast_fp16 = mul(x = p2c_att_3_cast_fp16, y = _inversed_331_y_0_to_fp16)[name = string("_inversed_331_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_1_cast_fp16 = add(x = _inversed_305_cast_fp16, y = _inversed_331_cast_fp16)[name = string("rel_att_1_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_3_cast_fp16 = add(x = attention_scores_1_cast_fp16, y = rel_att_1_cast_fp16)[name = string("attention_scores_3_cast_fp16")];
-            tensor<int32, [4]> var_336 = const()[name = string("op_336"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_5_cast_fp16 = reshape(shape = var_336, x = attention_scores_3_cast_fp16)[name = string("attention_scores_5_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_3_cast_fp16 = transpose(perm = p2c_att_3_perm_0, x = var_328_cast_fp16_cast_int16)[name = string("transpose_139")];
+            tensor<fp16, [16, 192, 192]> _inversed_331_cast_fp16 = mul(x = p2c_att_3_cast_fp16, y = _inversed_331_y_0_to_fp16)[name = string("_inversed_331_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_1_cast_fp16 = add(x = _inversed_305_cast_fp16, y = _inversed_331_cast_fp16)[name = string("rel_att_1_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_3_cast_fp16 = add(x = attention_scores_1_cast_fp16, y = rel_att_1_cast_fp16)[name = string("attention_scores_3_cast_fp16")];
+            tensor<int32, [4]> var_336 = const()[name = string("op_336"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_5_cast_fp16 = reshape(shape = var_336, x = attention_scores_3_cast_fp16)[name = string("attention_scores_5_cast_fp16")];
             string attention_mask_7_dtype_0 = const()[name = string("attention_mask_7_dtype_0"), val = string("bool")];
-            tensor<bool, [1, 1, 256, 256]> attention_mask_7 = cast(dtype = attention_mask_7_dtype_0, x = attention_mask_5)[name = string("cast_491")];
-            tensor<bool, [1, 1, 256, 256]> var_339 = logical_not(x = attention_mask_7)[name = string("op_339")];
+            tensor<bool, [1, 1, 192, 192]> attention_mask_7 = cast(dtype = attention_mask_7_dtype_0, x = attention_mask_5)[name = string("cast_491")];
+            tensor<bool, [1, 1, 192, 192]> var_339 = logical_not(x = attention_mask_7)[name = string("op_339")];
             fp16 var_64_to_fp16 = const()[name = string("op_64_to_fp16"), val = fp16(-inf)];
-            tensor<fp16, [1, 16, 256, 256]> input_9_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_5_cast_fp16, cond = var_339)[name = string("input_9_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_11_cast_fp16 = softmax(axis = var_56, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
-            tensor<int32, [3]> var_345 = const()[name = string("op_345"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_346_cast_fp16 = reshape(shape = var_345, x = input_11_cast_fp16)[name = string("op_346_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_9_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_5_cast_fp16, cond = var_339)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_11_cast_fp16 = softmax(axis = var_56, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<int32, [3]> var_345 = const()[name = string("op_345"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_346_cast_fp16 = reshape(shape = var_345, x = input_11_cast_fp16)[name = string("op_346_cast_fp16")];
             bool context_layer_1_transpose_x_0 = const()[name = string("context_layer_1_transpose_x_0"), val = bool(false)];
             bool context_layer_1_transpose_y_0 = const()[name = string("context_layer_1_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = var_346_cast_fp16, y = value_layer_1_cast_fp16)[name = string("context_layer_1_cast_fp16")];
-            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_351_cast_fp16 = reshape(shape = var_350, x = context_layer_1_cast_fp16)[name = string("op_351_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = var_346_cast_fp16, y = value_layer_1_cast_fp16)[name = string("context_layer_1_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_351_cast_fp16 = reshape(shape = var_350, x = context_layer_1_cast_fp16)[name = string("op_351_cast_fp16")];
             tensor<int32, [4]> var_352 = const()[name = string("op_352"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_357 = const()[name = string("op_357"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_353_cast_fp16 = transpose(perm = var_352, x = var_351_cast_fp16)[name = string("transpose_138")];
-            tensor<fp16, [1, 256, 1024]> input_13_cast_fp16 = reshape(shape = var_357, x = var_353_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<int32, [3]> var_357 = const()[name = string("op_357"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_353_cast_fp16 = transpose(perm = var_352, x = var_351_cast_fp16)[name = string("transpose_138")];
+            tensor<fp16, [1, 192, 1024]> input_13_cast_fp16 = reshape(shape = var_357, x = var_353_cast_fp16)[name = string("input_13_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271615424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273712640)))];
-            tensor<fp16, [1, 256, 1024]> linear_5_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_5_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_17_cast_fp16 = add(x = linear_5_cast_fp16, y = input_1_cast_fp16)[name = string("input_17_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_5_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_17_cast_fp16 = add(x = linear_5_cast_fp16, y = input_1_cast_fp16)[name = string("input_17_cast_fp16")];
             tensor<int32, [1]> input_19_axes_0 = const()[name = string("input_19_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273714752)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273716864)))];
             fp16 var_67_to_fp16 = const()[name = string("op_67_to_fp16"), val = fp16(0x1p-23)];
-            tensor<fp16, [1, 256, 1024]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_0_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273718976)))];
             tensor<fp16, [4096]> encoder_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282107648)))];
-            tensor<fp16, [1, 256, 4096]> linear_6_cast_fp16 = linear(bias = encoder_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_6_cast_fp16 = linear(bias = encoder_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
             string input_23_mode_0 = const()[name = string("input_23_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = string("input_23_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_0_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282115904)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290504576)))];
-            tensor<fp16, [1, 256, 1024]> linear_7_cast_fp16 = linear(bias = encoder_encoder_layer_0_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_output_dense_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_27_cast_fp16 = add(x = linear_7_cast_fp16, y = input_19_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_7_cast_fp16 = linear(bias = encoder_encoder_layer_0_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_output_dense_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_27_cast_fp16 = add(x = linear_7_cast_fp16, y = input_19_cast_fp16)[name = string("input_27_cast_fp16")];
             tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290506688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290508800)))];
-            tensor<fp16, [1, 256, 1024]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290510912)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292608128)))];
-            tensor<fp16, [1, 256, 1024]> linear_8_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_8_cast_fp16")];
-            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_23_cast_fp16 = reshape(shape = var_408, x = linear_8_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_8_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_23_cast_fp16 = reshape(shape = var_408, x = linear_8_cast_fp16)[name = string("x_23_cast_fp16")];
             tensor<int32, [4]> var_410 = const()[name = string("op_410"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_411_cast_fp16 = transpose(perm = var_410, x = x_23_cast_fp16)[name = string("transpose_137")];
-            tensor<fp16, [16, 256, 64]> query_layer_9_cast_fp16 = reshape(shape = var_415, x = var_411_cast_fp16)[name = string("query_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_411_cast_fp16 = transpose(perm = var_410, x = x_23_cast_fp16)[name = string("transpose_137")];
+            tensor<fp16, [16, 192, 64]> query_layer_9_cast_fp16 = reshape(shape = var_415, x = var_411_cast_fp16)[name = string("query_layer_9_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292610240)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294707456)))];
-            tensor<fp16, [1, 256, 1024]> linear_9_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_9_cast_fp16")];
-            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_27_cast_fp16 = reshape(shape = var_422, x = linear_9_cast_fp16)[name = string("x_27_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_9_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_27_cast_fp16 = reshape(shape = var_422, x = linear_9_cast_fp16)[name = string("x_27_cast_fp16")];
             tensor<int32, [4]> var_424 = const()[name = string("op_424"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_425_cast_fp16 = transpose(perm = var_424, x = x_27_cast_fp16)[name = string("transpose_136")];
-            tensor<fp16, [16, 256, 64]> key_layer_3_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = string("key_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_425_cast_fp16 = transpose(perm = var_424, x = x_27_cast_fp16)[name = string("transpose_136")];
+            tensor<fp16, [16, 192, 64]> key_layer_3_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = string("key_layer_3_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294709568)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296806784)))];
-            tensor<fp16, [1, 256, 1024]> linear_10_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_10_cast_fp16")];
-            tensor<int32, [4]> var_436 = const()[name = string("op_436"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_31_cast_fp16 = reshape(shape = var_436, x = linear_10_cast_fp16)[name = string("x_31_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_10_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<int32, [4]> var_436 = const()[name = string("op_436"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_31_cast_fp16 = reshape(shape = var_436, x = linear_10_cast_fp16)[name = string("x_31_cast_fp16")];
             tensor<int32, [4]> var_438 = const()[name = string("op_438"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_443 = const()[name = string("op_443"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_439_cast_fp16 = transpose(perm = var_438, x = x_31_cast_fp16)[name = string("transpose_135")];
-            tensor<fp16, [16, 256, 64]> value_layer_3_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = string("value_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_443 = const()[name = string("op_443"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_439_cast_fp16 = transpose(perm = var_438, x = x_31_cast_fp16)[name = string("transpose_135")];
+            tensor<fp16, [16, 192, 64]> value_layer_3_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = string("value_layer_3_cast_fp16")];
             tensor<int32, [3]> var_453_perm_0 = const()[name = string("op_453_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_455_y_0_to_fp16 = const()[name = string("_inversed_455_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_453_cast_fp16 = transpose(perm = var_453_perm_0, x = key_layer_3_cast_fp16)[name = string("transpose_134")];
-            tensor<fp16, [16, 64, 256]> _inversed_455_cast_fp16 = mul(x = var_453_cast_fp16, y = _inversed_455_y_0_to_fp16)[name = string("_inversed_455_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_453_cast_fp16 = transpose(perm = var_453_perm_0, x = key_layer_3_cast_fp16)[name = string("transpose_134")];
+            tensor<fp16, [16, 64, 192]> _inversed_455_cast_fp16 = mul(x = var_453_cast_fp16, y = _inversed_455_y_0_to_fp16)[name = string("_inversed_455_cast_fp16")];
             bool attention_scores_7_transpose_x_0 = const()[name = string("attention_scores_7_transpose_x_0"), val = bool(false)];
             bool attention_scores_7_transpose_y_0 = const()[name = string("attention_scores_7_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = query_layer_9_cast_fp16, y = _inversed_455_cast_fp16)[name = string("attention_scores_7_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = query_layer_9_cast_fp16, y = _inversed_455_cast_fp16)[name = string("attention_scores_7_cast_fp16")];
             bool c2p_att_5_transpose_x_0 = const()[name = string("c2p_att_5_transpose_x_0"), val = bool(false)];
             bool c2p_att_5_transpose_y_0 = const()[name = string("c2p_att_5_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_511_to_fp16 = const()[name = string("op_511_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296808896)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_5_cast_fp16 = matmul(transpose_x = c2p_att_5_transpose_x_0, transpose_y = c2p_att_5_transpose_y_0, x = query_layer_9_cast_fp16, y = var_511_to_fp16)[name = string("c2p_att_5_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_5_cast_fp16 = matmul(transpose_x = c2p_att_5_transpose_x_0, transpose_y = c2p_att_5_transpose_y_0, x = query_layer_9_cast_fp16, y = var_511_to_fp16)[name = string("c2p_att_5_cast_fp16")];
             bool c2p_att_7_validate_indices_0 = const()[name = string("c2p_att_7_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_7_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_7_validate_indices_0, x = c2p_att_5_cast_fp16)[name = string("c2p_att_7_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_7_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_7_validate_indices_0, x = c2p_att_5_cast_fp16)[name = string("c2p_att_7_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_524_y_0_to_fp16 = const()[name = string("_inversed_524_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_524_cast_fp16 = mul(x = c2p_att_7_cast_fp16_cast_int16, y = _inversed_524_y_0_to_fp16)[name = string("_inversed_524_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_524_cast_fp16 = mul(x = c2p_att_7_cast_fp16_cast_int16, y = _inversed_524_y_0_to_fp16)[name = string("_inversed_524_cast_fp16")];
             bool p2c_att_5_transpose_x_0 = const()[name = string("p2c_att_5_transpose_x_0"), val = bool(false)];
             bool p2c_att_5_transpose_y_0 = const()[name = string("p2c_att_5_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_539_to_fp16 = const()[name = string("op_539_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297857536)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_5_cast_fp16 = matmul(transpose_x = p2c_att_5_transpose_x_0, transpose_y = p2c_att_5_transpose_y_0, x = key_layer_3_cast_fp16, y = var_539_to_fp16)[name = string("p2c_att_5_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_5_cast_fp16 = matmul(transpose_x = p2c_att_5_transpose_x_0, transpose_y = p2c_att_5_transpose_y_0, x = key_layer_3_cast_fp16, y = var_539_to_fp16)[name = string("p2c_att_5_cast_fp16")];
             bool var_547_validate_indices_0 = const()[name = string("op_547_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_547_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_547_validate_indices_0, x = p2c_att_5_cast_fp16)[name = string("op_547_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_547_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_547_validate_indices_0, x = p2c_att_5_cast_fp16)[name = string("op_547_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_7_perm_0 = const()[name = string("p2c_att_7_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_550_y_0_to_fp16 = const()[name = string("_inversed_550_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_7_cast_fp16 = transpose(perm = p2c_att_7_perm_0, x = var_547_cast_fp16_cast_int16)[name = string("transpose_133")];
-            tensor<fp16, [16, 256, 256]> _inversed_550_cast_fp16 = mul(x = p2c_att_7_cast_fp16, y = _inversed_550_y_0_to_fp16)[name = string("_inversed_550_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_3_cast_fp16 = add(x = _inversed_524_cast_fp16, y = _inversed_550_cast_fp16)[name = string("rel_att_3_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_9_cast_fp16 = add(x = attention_scores_7_cast_fp16, y = rel_att_3_cast_fp16)[name = string("attention_scores_9_cast_fp16")];
-            tensor<int32, [4]> var_555 = const()[name = string("op_555"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_11_cast_fp16 = reshape(shape = var_555, x = attention_scores_9_cast_fp16)[name = string("attention_scores_11_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_33_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_11_cast_fp16, cond = var_339)[name = string("input_33_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_35_cast_fp16 = softmax(axis = var_56, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
-            tensor<int32, [3]> var_564 = const()[name = string("op_564"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_565_cast_fp16 = reshape(shape = var_564, x = input_35_cast_fp16)[name = string("op_565_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_7_cast_fp16 = transpose(perm = p2c_att_7_perm_0, x = var_547_cast_fp16_cast_int16)[name = string("transpose_133")];
+            tensor<fp16, [16, 192, 192]> _inversed_550_cast_fp16 = mul(x = p2c_att_7_cast_fp16, y = _inversed_550_y_0_to_fp16)[name = string("_inversed_550_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_3_cast_fp16 = add(x = _inversed_524_cast_fp16, y = _inversed_550_cast_fp16)[name = string("rel_att_3_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_9_cast_fp16 = add(x = attention_scores_7_cast_fp16, y = rel_att_3_cast_fp16)[name = string("attention_scores_9_cast_fp16")];
+            tensor<int32, [4]> var_555 = const()[name = string("op_555"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_11_cast_fp16 = reshape(shape = var_555, x = attention_scores_9_cast_fp16)[name = string("attention_scores_11_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_33_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_11_cast_fp16, cond = var_339)[name = string("input_33_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_35_cast_fp16 = softmax(axis = var_56, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
+            tensor<int32, [3]> var_564 = const()[name = string("op_564"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_565_cast_fp16 = reshape(shape = var_564, x = input_35_cast_fp16)[name = string("op_565_cast_fp16")];
             bool context_layer_5_transpose_x_0 = const()[name = string("context_layer_5_transpose_x_0"), val = bool(false)];
             bool context_layer_5_transpose_y_0 = const()[name = string("context_layer_5_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = var_565_cast_fp16, y = value_layer_3_cast_fp16)[name = string("context_layer_5_cast_fp16")];
-            tensor<int32, [4]> var_569 = const()[name = string("op_569"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_570_cast_fp16 = reshape(shape = var_569, x = context_layer_5_cast_fp16)[name = string("op_570_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = var_565_cast_fp16, y = value_layer_3_cast_fp16)[name = string("context_layer_5_cast_fp16")];
+            tensor<int32, [4]> var_569 = const()[name = string("op_569"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_570_cast_fp16 = reshape(shape = var_569, x = context_layer_5_cast_fp16)[name = string("op_570_cast_fp16")];
             tensor<int32, [4]> var_571 = const()[name = string("op_571"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_576 = const()[name = string("op_576"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_572_cast_fp16 = transpose(perm = var_571, x = var_570_cast_fp16)[name = string("transpose_132")];
-            tensor<fp16, [1, 256, 1024]> input_37_cast_fp16 = reshape(shape = var_576, x = var_572_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [3]> var_576 = const()[name = string("op_576"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_572_cast_fp16 = transpose(perm = var_571, x = var_570_cast_fp16)[name = string("transpose_132")];
+            tensor<fp16, [1, 192, 1024]> input_37_cast_fp16 = reshape(shape = var_576, x = var_572_cast_fp16)[name = string("input_37_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298906176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301003392)))];
-            tensor<fp16, [1, 256, 1024]> linear_13_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_13_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_41_cast_fp16 = add(x = linear_13_cast_fp16, y = input_29_cast_fp16)[name = string("input_41_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_13_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_41_cast_fp16 = add(x = linear_13_cast_fp16, y = input_29_cast_fp16)[name = string("input_41_cast_fp16")];
             tensor<int32, [1]> input_43_axes_0 = const()[name = string("input_43_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301005504)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301007616)))];
-            tensor<fp16, [1, 256, 1024]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_1_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301009728)))];
             tensor<fp16, [4096]> encoder_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309398400)))];
-            tensor<fp16, [1, 256, 4096]> linear_14_cast_fp16 = linear(bias = encoder_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_14_cast_fp16 = linear(bias = encoder_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_14_cast_fp16")];
             string input_47_mode_0 = const()[name = string("input_47_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = string("input_47_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_1_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309406656)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317795328)))];
-            tensor<fp16, [1, 256, 1024]> linear_15_cast_fp16 = linear(bias = encoder_encoder_layer_1_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_output_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_15_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_51_cast_fp16 = add(x = linear_15_cast_fp16, y = input_43_cast_fp16)[name = string("input_51_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_15_cast_fp16 = linear(bias = encoder_encoder_layer_1_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_output_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_51_cast_fp16 = add(x = linear_15_cast_fp16, y = input_43_cast_fp16)[name = string("input_51_cast_fp16")];
             tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317797440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317799552)))];
-            tensor<fp16, [1, 256, 1024]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317801664)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319898880)))];
-            tensor<fp16, [1, 256, 1024]> linear_16_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_16_cast_fp16")];
-            tensor<int32, [4]> var_627 = const()[name = string("op_627"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_43_cast_fp16 = reshape(shape = var_627, x = linear_16_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_16_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<int32, [4]> var_627 = const()[name = string("op_627"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_43_cast_fp16 = reshape(shape = var_627, x = linear_16_cast_fp16)[name = string("x_43_cast_fp16")];
             tensor<int32, [4]> var_629 = const()[name = string("op_629"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_634 = const()[name = string("op_634"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_630_cast_fp16 = transpose(perm = var_629, x = x_43_cast_fp16)[name = string("transpose_131")];
-            tensor<fp16, [16, 256, 64]> query_layer_15_cast_fp16 = reshape(shape = var_634, x = var_630_cast_fp16)[name = string("query_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_634 = const()[name = string("op_634"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_630_cast_fp16 = transpose(perm = var_629, x = x_43_cast_fp16)[name = string("transpose_131")];
+            tensor<fp16, [16, 192, 64]> query_layer_15_cast_fp16 = reshape(shape = var_634, x = var_630_cast_fp16)[name = string("query_layer_15_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319900992)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321998208)))];
-            tensor<fp16, [1, 256, 1024]> linear_17_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_17_cast_fp16")];
-            tensor<int32, [4]> var_641 = const()[name = string("op_641"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_47_cast_fp16 = reshape(shape = var_641, x = linear_17_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_17_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<int32, [4]> var_641 = const()[name = string("op_641"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_47_cast_fp16 = reshape(shape = var_641, x = linear_17_cast_fp16)[name = string("x_47_cast_fp16")];
             tensor<int32, [4]> var_643 = const()[name = string("op_643"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_648 = const()[name = string("op_648"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_644_cast_fp16 = transpose(perm = var_643, x = x_47_cast_fp16)[name = string("transpose_130")];
-            tensor<fp16, [16, 256, 64]> key_layer_5_cast_fp16 = reshape(shape = var_648, x = var_644_cast_fp16)[name = string("key_layer_5_cast_fp16")];
+            tensor<int32, [3]> var_648 = const()[name = string("op_648"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_644_cast_fp16 = transpose(perm = var_643, x = x_47_cast_fp16)[name = string("transpose_130")];
+            tensor<fp16, [16, 192, 64]> key_layer_5_cast_fp16 = reshape(shape = var_648, x = var_644_cast_fp16)[name = string("key_layer_5_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322000320)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324097536)))];
-            tensor<fp16, [1, 256, 1024]> linear_18_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_18_cast_fp16")];
-            tensor<int32, [4]> var_655 = const()[name = string("op_655"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_51_cast_fp16 = reshape(shape = var_655, x = linear_18_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_18_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_18_cast_fp16")];
+            tensor<int32, [4]> var_655 = const()[name = string("op_655"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_51_cast_fp16 = reshape(shape = var_655, x = linear_18_cast_fp16)[name = string("x_51_cast_fp16")];
             tensor<int32, [4]> var_657 = const()[name = string("op_657"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_662 = const()[name = string("op_662"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_658_cast_fp16 = transpose(perm = var_657, x = x_51_cast_fp16)[name = string("transpose_129")];
-            tensor<fp16, [16, 256, 64]> value_layer_5_cast_fp16 = reshape(shape = var_662, x = var_658_cast_fp16)[name = string("value_layer_5_cast_fp16")];
+            tensor<int32, [3]> var_662 = const()[name = string("op_662"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_658_cast_fp16 = transpose(perm = var_657, x = x_51_cast_fp16)[name = string("transpose_129")];
+            tensor<fp16, [16, 192, 64]> value_layer_5_cast_fp16 = reshape(shape = var_662, x = var_658_cast_fp16)[name = string("value_layer_5_cast_fp16")];
             tensor<int32, [3]> var_672_perm_0 = const()[name = string("op_672_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_674_y_0_to_fp16 = const()[name = string("_inversed_674_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_672_cast_fp16 = transpose(perm = var_672_perm_0, x = key_layer_5_cast_fp16)[name = string("transpose_128")];
-            tensor<fp16, [16, 64, 256]> _inversed_674_cast_fp16 = mul(x = var_672_cast_fp16, y = _inversed_674_y_0_to_fp16)[name = string("_inversed_674_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_672_cast_fp16 = transpose(perm = var_672_perm_0, x = key_layer_5_cast_fp16)[name = string("transpose_128")];
+            tensor<fp16, [16, 64, 192]> _inversed_674_cast_fp16 = mul(x = var_672_cast_fp16, y = _inversed_674_y_0_to_fp16)[name = string("_inversed_674_cast_fp16")];
             bool attention_scores_13_transpose_x_0 = const()[name = string("attention_scores_13_transpose_x_0"), val = bool(false)];
             bool attention_scores_13_transpose_y_0 = const()[name = string("attention_scores_13_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = query_layer_15_cast_fp16, y = _inversed_674_cast_fp16)[name = string("attention_scores_13_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = query_layer_15_cast_fp16, y = _inversed_674_cast_fp16)[name = string("attention_scores_13_cast_fp16")];
             bool c2p_att_9_transpose_x_0 = const()[name = string("c2p_att_9_transpose_x_0"), val = bool(false)];
             bool c2p_att_9_transpose_y_0 = const()[name = string("c2p_att_9_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_730_to_fp16 = const()[name = string("op_730_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324099648)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_9_cast_fp16 = matmul(transpose_x = c2p_att_9_transpose_x_0, transpose_y = c2p_att_9_transpose_y_0, x = query_layer_15_cast_fp16, y = var_730_to_fp16)[name = string("c2p_att_9_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_9_cast_fp16 = matmul(transpose_x = c2p_att_9_transpose_x_0, transpose_y = c2p_att_9_transpose_y_0, x = query_layer_15_cast_fp16, y = var_730_to_fp16)[name = string("c2p_att_9_cast_fp16")];
             bool c2p_att_11_validate_indices_0 = const()[name = string("c2p_att_11_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_11_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_11_validate_indices_0, x = c2p_att_9_cast_fp16)[name = string("c2p_att_11_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_11_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_11_validate_indices_0, x = c2p_att_9_cast_fp16)[name = string("c2p_att_11_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_743_y_0_to_fp16 = const()[name = string("_inversed_743_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_743_cast_fp16 = mul(x = c2p_att_11_cast_fp16_cast_int16, y = _inversed_743_y_0_to_fp16)[name = string("_inversed_743_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_743_cast_fp16 = mul(x = c2p_att_11_cast_fp16_cast_int16, y = _inversed_743_y_0_to_fp16)[name = string("_inversed_743_cast_fp16")];
             bool p2c_att_9_transpose_x_0 = const()[name = string("p2c_att_9_transpose_x_0"), val = bool(false)];
             bool p2c_att_9_transpose_y_0 = const()[name = string("p2c_att_9_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_758_to_fp16 = const()[name = string("op_758_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325148288)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_9_cast_fp16 = matmul(transpose_x = p2c_att_9_transpose_x_0, transpose_y = p2c_att_9_transpose_y_0, x = key_layer_5_cast_fp16, y = var_758_to_fp16)[name = string("p2c_att_9_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_9_cast_fp16 = matmul(transpose_x = p2c_att_9_transpose_x_0, transpose_y = p2c_att_9_transpose_y_0, x = key_layer_5_cast_fp16, y = var_758_to_fp16)[name = string("p2c_att_9_cast_fp16")];
             bool var_766_validate_indices_0 = const()[name = string("op_766_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_766_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_766_validate_indices_0, x = p2c_att_9_cast_fp16)[name = string("op_766_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_766_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_766_validate_indices_0, x = p2c_att_9_cast_fp16)[name = string("op_766_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_11_perm_0 = const()[name = string("p2c_att_11_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_769_y_0_to_fp16 = const()[name = string("_inversed_769_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_11_cast_fp16 = transpose(perm = p2c_att_11_perm_0, x = var_766_cast_fp16_cast_int16)[name = string("transpose_127")];
-            tensor<fp16, [16, 256, 256]> _inversed_769_cast_fp16 = mul(x = p2c_att_11_cast_fp16, y = _inversed_769_y_0_to_fp16)[name = string("_inversed_769_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_5_cast_fp16 = add(x = _inversed_743_cast_fp16, y = _inversed_769_cast_fp16)[name = string("rel_att_5_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = rel_att_5_cast_fp16)[name = string("attention_scores_15_cast_fp16")];
-            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_17_cast_fp16 = reshape(shape = var_774, x = attention_scores_15_cast_fp16)[name = string("attention_scores_17_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_57_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_17_cast_fp16, cond = var_339)[name = string("input_57_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_59_cast_fp16 = softmax(axis = var_56, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
-            tensor<int32, [3]> var_783 = const()[name = string("op_783"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_784_cast_fp16 = reshape(shape = var_783, x = input_59_cast_fp16)[name = string("op_784_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_11_cast_fp16 = transpose(perm = p2c_att_11_perm_0, x = var_766_cast_fp16_cast_int16)[name = string("transpose_127")];
+            tensor<fp16, [16, 192, 192]> _inversed_769_cast_fp16 = mul(x = p2c_att_11_cast_fp16, y = _inversed_769_y_0_to_fp16)[name = string("_inversed_769_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_5_cast_fp16 = add(x = _inversed_743_cast_fp16, y = _inversed_769_cast_fp16)[name = string("rel_att_5_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = rel_att_5_cast_fp16)[name = string("attention_scores_15_cast_fp16")];
+            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_17_cast_fp16 = reshape(shape = var_774, x = attention_scores_15_cast_fp16)[name = string("attention_scores_17_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_57_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_17_cast_fp16, cond = var_339)[name = string("input_57_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_59_cast_fp16 = softmax(axis = var_56, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            tensor<int32, [3]> var_783 = const()[name = string("op_783"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_784_cast_fp16 = reshape(shape = var_783, x = input_59_cast_fp16)[name = string("op_784_cast_fp16")];
             bool context_layer_9_transpose_x_0 = const()[name = string("context_layer_9_transpose_x_0"), val = bool(false)];
             bool context_layer_9_transpose_y_0 = const()[name = string("context_layer_9_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = var_784_cast_fp16, y = value_layer_5_cast_fp16)[name = string("context_layer_9_cast_fp16")];
-            tensor<int32, [4]> var_788 = const()[name = string("op_788"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_789_cast_fp16 = reshape(shape = var_788, x = context_layer_9_cast_fp16)[name = string("op_789_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = var_784_cast_fp16, y = value_layer_5_cast_fp16)[name = string("context_layer_9_cast_fp16")];
+            tensor<int32, [4]> var_788 = const()[name = string("op_788"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_789_cast_fp16 = reshape(shape = var_788, x = context_layer_9_cast_fp16)[name = string("op_789_cast_fp16")];
             tensor<int32, [4]> var_790 = const()[name = string("op_790"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_795 = const()[name = string("op_795"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_791_cast_fp16 = transpose(perm = var_790, x = var_789_cast_fp16)[name = string("transpose_126")];
-            tensor<fp16, [1, 256, 1024]> input_61_cast_fp16 = reshape(shape = var_795, x = var_791_cast_fp16)[name = string("input_61_cast_fp16")];
+            tensor<int32, [3]> var_795 = const()[name = string("op_795"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_791_cast_fp16 = transpose(perm = var_790, x = var_789_cast_fp16)[name = string("transpose_126")];
+            tensor<fp16, [1, 192, 1024]> input_61_cast_fp16 = reshape(shape = var_795, x = var_791_cast_fp16)[name = string("input_61_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326196928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328294144)))];
-            tensor<fp16, [1, 256, 1024]> linear_21_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_21_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_53_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_21_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_53_cast_fp16)[name = string("input_65_cast_fp16")];
             tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328296256)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328298368)))];
-            tensor<fp16, [1, 256, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_2_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328300480)))];
             tensor<fp16, [4096]> encoder_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336689152)))];
-            tensor<fp16, [1, 256, 4096]> linear_22_cast_fp16 = linear(bias = encoder_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_22_cast_fp16 = linear(bias = encoder_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_22_cast_fp16")];
             string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = string("input_71_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = string("input_71_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_2_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336697408)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345086080)))];
-            tensor<fp16, [1, 256, 1024]> linear_23_cast_fp16 = linear(bias = encoder_encoder_layer_2_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_output_dense_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_23_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_67_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_23_cast_fp16 = linear(bias = encoder_encoder_layer_2_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_output_dense_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_67_cast_fp16)[name = string("input_75_cast_fp16")];
             tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345088192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345090304)))];
-            tensor<fp16, [1, 256, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092416)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347189632)))];
-            tensor<fp16, [1, 256, 1024]> linear_24_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_24_cast_fp16")];
-            tensor<int32, [4]> var_846 = const()[name = string("op_846"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_63_cast_fp16 = reshape(shape = var_846, x = linear_24_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_24_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_24_cast_fp16")];
+            tensor<int32, [4]> var_846 = const()[name = string("op_846"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_63_cast_fp16 = reshape(shape = var_846, x = linear_24_cast_fp16)[name = string("x_63_cast_fp16")];
             tensor<int32, [4]> var_848 = const()[name = string("op_848"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_853 = const()[name = string("op_853"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_849_cast_fp16 = transpose(perm = var_848, x = x_63_cast_fp16)[name = string("transpose_125")];
-            tensor<fp16, [16, 256, 64]> query_layer_21_cast_fp16 = reshape(shape = var_853, x = var_849_cast_fp16)[name = string("query_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_853 = const()[name = string("op_853"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_849_cast_fp16 = transpose(perm = var_848, x = x_63_cast_fp16)[name = string("transpose_125")];
+            tensor<fp16, [16, 192, 64]> query_layer_21_cast_fp16 = reshape(shape = var_853, x = var_849_cast_fp16)[name = string("query_layer_21_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347191744)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349288960)))];
-            tensor<fp16, [1, 256, 1024]> linear_25_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_25_cast_fp16")];
-            tensor<int32, [4]> var_860 = const()[name = string("op_860"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_67_cast_fp16 = reshape(shape = var_860, x = linear_25_cast_fp16)[name = string("x_67_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_25_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<int32, [4]> var_860 = const()[name = string("op_860"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_67_cast_fp16 = reshape(shape = var_860, x = linear_25_cast_fp16)[name = string("x_67_cast_fp16")];
             tensor<int32, [4]> var_862 = const()[name = string("op_862"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_867 = const()[name = string("op_867"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_863_cast_fp16 = transpose(perm = var_862, x = x_67_cast_fp16)[name = string("transpose_124")];
-            tensor<fp16, [16, 256, 64]> key_layer_7_cast_fp16 = reshape(shape = var_867, x = var_863_cast_fp16)[name = string("key_layer_7_cast_fp16")];
+            tensor<int32, [3]> var_867 = const()[name = string("op_867"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_863_cast_fp16 = transpose(perm = var_862, x = x_67_cast_fp16)[name = string("transpose_124")];
+            tensor<fp16, [16, 192, 64]> key_layer_7_cast_fp16 = reshape(shape = var_867, x = var_863_cast_fp16)[name = string("key_layer_7_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349291072)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351388288)))];
-            tensor<fp16, [1, 256, 1024]> linear_26_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_26_cast_fp16")];
-            tensor<int32, [4]> var_874 = const()[name = string("op_874"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_71_cast_fp16 = reshape(shape = var_874, x = linear_26_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_26_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<int32, [4]> var_874 = const()[name = string("op_874"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_71_cast_fp16 = reshape(shape = var_874, x = linear_26_cast_fp16)[name = string("x_71_cast_fp16")];
             tensor<int32, [4]> var_876 = const()[name = string("op_876"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_881 = const()[name = string("op_881"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_877_cast_fp16 = transpose(perm = var_876, x = x_71_cast_fp16)[name = string("transpose_123")];
-            tensor<fp16, [16, 256, 64]> value_layer_7_cast_fp16 = reshape(shape = var_881, x = var_877_cast_fp16)[name = string("value_layer_7_cast_fp16")];
+            tensor<int32, [3]> var_881 = const()[name = string("op_881"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_877_cast_fp16 = transpose(perm = var_876, x = x_71_cast_fp16)[name = string("transpose_123")];
+            tensor<fp16, [16, 192, 64]> value_layer_7_cast_fp16 = reshape(shape = var_881, x = var_877_cast_fp16)[name = string("value_layer_7_cast_fp16")];
             tensor<int32, [3]> var_891_perm_0 = const()[name = string("op_891_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_893_y_0_to_fp16 = const()[name = string("_inversed_893_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_891_cast_fp16 = transpose(perm = var_891_perm_0, x = key_layer_7_cast_fp16)[name = string("transpose_122")];
-            tensor<fp16, [16, 64, 256]> _inversed_893_cast_fp16 = mul(x = var_891_cast_fp16, y = _inversed_893_y_0_to_fp16)[name = string("_inversed_893_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_891_cast_fp16 = transpose(perm = var_891_perm_0, x = key_layer_7_cast_fp16)[name = string("transpose_122")];
+            tensor<fp16, [16, 64, 192]> _inversed_893_cast_fp16 = mul(x = var_891_cast_fp16, y = _inversed_893_y_0_to_fp16)[name = string("_inversed_893_cast_fp16")];
             bool attention_scores_19_transpose_x_0 = const()[name = string("attention_scores_19_transpose_x_0"), val = bool(false)];
             bool attention_scores_19_transpose_y_0 = const()[name = string("attention_scores_19_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = query_layer_21_cast_fp16, y = _inversed_893_cast_fp16)[name = string("attention_scores_19_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = query_layer_21_cast_fp16, y = _inversed_893_cast_fp16)[name = string("attention_scores_19_cast_fp16")];
             bool c2p_att_13_transpose_x_0 = const()[name = string("c2p_att_13_transpose_x_0"), val = bool(false)];
             bool c2p_att_13_transpose_y_0 = const()[name = string("c2p_att_13_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_949_to_fp16 = const()[name = string("op_949_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351390400)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_13_cast_fp16 = matmul(transpose_x = c2p_att_13_transpose_x_0, transpose_y = c2p_att_13_transpose_y_0, x = query_layer_21_cast_fp16, y = var_949_to_fp16)[name = string("c2p_att_13_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_13_cast_fp16 = matmul(transpose_x = c2p_att_13_transpose_x_0, transpose_y = c2p_att_13_transpose_y_0, x = query_layer_21_cast_fp16, y = var_949_to_fp16)[name = string("c2p_att_13_cast_fp16")];
             bool c2p_att_15_validate_indices_0 = const()[name = string("c2p_att_15_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_15_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_15_validate_indices_0, x = c2p_att_13_cast_fp16)[name = string("c2p_att_15_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_15_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_15_validate_indices_0, x = c2p_att_13_cast_fp16)[name = string("c2p_att_15_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_962_y_0_to_fp16 = const()[name = string("_inversed_962_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_962_cast_fp16 = mul(x = c2p_att_15_cast_fp16_cast_int16, y = _inversed_962_y_0_to_fp16)[name = string("_inversed_962_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_962_cast_fp16 = mul(x = c2p_att_15_cast_fp16_cast_int16, y = _inversed_962_y_0_to_fp16)[name = string("_inversed_962_cast_fp16")];
             bool p2c_att_13_transpose_x_0 = const()[name = string("p2c_att_13_transpose_x_0"), val = bool(false)];
             bool p2c_att_13_transpose_y_0 = const()[name = string("p2c_att_13_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_977_to_fp16 = const()[name = string("op_977_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352439040)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_13_cast_fp16 = matmul(transpose_x = p2c_att_13_transpose_x_0, transpose_y = p2c_att_13_transpose_y_0, x = key_layer_7_cast_fp16, y = var_977_to_fp16)[name = string("p2c_att_13_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_13_cast_fp16 = matmul(transpose_x = p2c_att_13_transpose_x_0, transpose_y = p2c_att_13_transpose_y_0, x = key_layer_7_cast_fp16, y = var_977_to_fp16)[name = string("p2c_att_13_cast_fp16")];
             bool var_985_validate_indices_0 = const()[name = string("op_985_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_985_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_985_validate_indices_0, x = p2c_att_13_cast_fp16)[name = string("op_985_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_985_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_985_validate_indices_0, x = p2c_att_13_cast_fp16)[name = string("op_985_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_15_perm_0 = const()[name = string("p2c_att_15_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_988_y_0_to_fp16 = const()[name = string("_inversed_988_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_15_cast_fp16 = transpose(perm = p2c_att_15_perm_0, x = var_985_cast_fp16_cast_int16)[name = string("transpose_121")];
-            tensor<fp16, [16, 256, 256]> _inversed_988_cast_fp16 = mul(x = p2c_att_15_cast_fp16, y = _inversed_988_y_0_to_fp16)[name = string("_inversed_988_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_7_cast_fp16 = add(x = _inversed_962_cast_fp16, y = _inversed_988_cast_fp16)[name = string("rel_att_7_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_21_cast_fp16 = add(x = attention_scores_19_cast_fp16, y = rel_att_7_cast_fp16)[name = string("attention_scores_21_cast_fp16")];
-            tensor<int32, [4]> var_993 = const()[name = string("op_993"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_23_cast_fp16 = reshape(shape = var_993, x = attention_scores_21_cast_fp16)[name = string("attention_scores_23_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_81_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_23_cast_fp16, cond = var_339)[name = string("input_81_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_83_cast_fp16 = softmax(axis = var_56, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
-            tensor<int32, [3]> var_1002 = const()[name = string("op_1002"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_1003_cast_fp16 = reshape(shape = var_1002, x = input_83_cast_fp16)[name = string("op_1003_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_15_cast_fp16 = transpose(perm = p2c_att_15_perm_0, x = var_985_cast_fp16_cast_int16)[name = string("transpose_121")];
+            tensor<fp16, [16, 192, 192]> _inversed_988_cast_fp16 = mul(x = p2c_att_15_cast_fp16, y = _inversed_988_y_0_to_fp16)[name = string("_inversed_988_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_7_cast_fp16 = add(x = _inversed_962_cast_fp16, y = _inversed_988_cast_fp16)[name = string("rel_att_7_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_21_cast_fp16 = add(x = attention_scores_19_cast_fp16, y = rel_att_7_cast_fp16)[name = string("attention_scores_21_cast_fp16")];
+            tensor<int32, [4]> var_993 = const()[name = string("op_993"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_23_cast_fp16 = reshape(shape = var_993, x = attention_scores_21_cast_fp16)[name = string("attention_scores_23_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_81_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_23_cast_fp16, cond = var_339)[name = string("input_81_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_83_cast_fp16 = softmax(axis = var_56, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<int32, [3]> var_1002 = const()[name = string("op_1002"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_1003_cast_fp16 = reshape(shape = var_1002, x = input_83_cast_fp16)[name = string("op_1003_cast_fp16")];
             bool context_layer_13_transpose_x_0 = const()[name = string("context_layer_13_transpose_x_0"), val = bool(false)];
             bool context_layer_13_transpose_y_0 = const()[name = string("context_layer_13_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = var_1003_cast_fp16, y = value_layer_7_cast_fp16)[name = string("context_layer_13_cast_fp16")];
-            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1008_cast_fp16 = reshape(shape = var_1007, x = context_layer_13_cast_fp16)[name = string("op_1008_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = var_1003_cast_fp16, y = value_layer_7_cast_fp16)[name = string("context_layer_13_cast_fp16")];
+            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1008_cast_fp16 = reshape(shape = var_1007, x = context_layer_13_cast_fp16)[name = string("op_1008_cast_fp16")];
             tensor<int32, [4]> var_1009 = const()[name = string("op_1009"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_1010_cast_fp16 = transpose(perm = var_1009, x = var_1008_cast_fp16)[name = string("transpose_120")];
-            tensor<fp16, [1, 256, 1024]> input_85_cast_fp16 = reshape(shape = var_1014, x = var_1010_cast_fp16)[name = string("input_85_cast_fp16")];
+            tensor<int32, [3]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_1010_cast_fp16 = transpose(perm = var_1009, x = var_1008_cast_fp16)[name = string("transpose_120")];
+            tensor<fp16, [1, 192, 1024]> input_85_cast_fp16 = reshape(shape = var_1014, x = var_1010_cast_fp16)[name = string("input_85_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353487680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355584896)))];
-            tensor<fp16, [1, 256, 1024]> linear_29_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_29_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_89_cast_fp16 = add(x = linear_29_cast_fp16, y = input_77_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_29_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_29_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_89_cast_fp16 = add(x = linear_29_cast_fp16, y = input_77_cast_fp16)[name = string("input_89_cast_fp16")];
             tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355587008)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355589120)))];
-            tensor<fp16, [1, 256, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_3_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355591232)))];
             tensor<fp16, [4096]> encoder_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363979904)))];
-            tensor<fp16, [1, 256, 4096]> linear_30_cast_fp16 = linear(bias = encoder_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_30_cast_fp16 = linear(bias = encoder_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_30_cast_fp16")];
             string input_95_mode_0 = const()[name = string("input_95_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = string("input_95_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_3_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363988160)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372376832)))];
-            tensor<fp16, [1, 256, 1024]> linear_31_cast_fp16 = linear(bias = encoder_encoder_layer_3_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_output_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_31_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_99_cast_fp16 = add(x = linear_31_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_31_cast_fp16 = linear(bias = encoder_encoder_layer_3_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_output_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_99_cast_fp16 = add(x = linear_31_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
             tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372378944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372381056)))];
-            tensor<fp16, [1, 256, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372383168)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374480384)))];
-            tensor<fp16, [1, 256, 1024]> linear_32_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_32_cast_fp16")];
-            tensor<int32, [4]> var_1065 = const()[name = string("op_1065"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_83_cast_fp16 = reshape(shape = var_1065, x = linear_32_cast_fp16)[name = string("x_83_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_32_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1065 = const()[name = string("op_1065"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_83_cast_fp16 = reshape(shape = var_1065, x = linear_32_cast_fp16)[name = string("x_83_cast_fp16")];
             tensor<int32, [4]> var_1067 = const()[name = string("op_1067"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1068_cast_fp16 = transpose(perm = var_1067, x = x_83_cast_fp16)[name = string("transpose_119")];
-            tensor<fp16, [16, 256, 64]> query_layer_27_cast_fp16 = reshape(shape = var_1072, x = var_1068_cast_fp16)[name = string("query_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1068_cast_fp16 = transpose(perm = var_1067, x = x_83_cast_fp16)[name = string("transpose_119")];
+            tensor<fp16, [16, 192, 64]> query_layer_27_cast_fp16 = reshape(shape = var_1072, x = var_1068_cast_fp16)[name = string("query_layer_27_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374482496)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376579712)))];
-            tensor<fp16, [1, 256, 1024]> linear_33_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")];
-            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_87_cast_fp16 = reshape(shape = var_1079, x = linear_33_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_33_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")];
+            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_87_cast_fp16 = reshape(shape = var_1079, x = linear_33_cast_fp16)[name = string("x_87_cast_fp16")];
             tensor<int32, [4]> var_1081 = const()[name = string("op_1081"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1082_cast_fp16 = transpose(perm = var_1081, x = x_87_cast_fp16)[name = string("transpose_118")];
-            tensor<fp16, [16, 256, 64]> key_layer_9_cast_fp16 = reshape(shape = var_1086, x = var_1082_cast_fp16)[name = string("key_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1082_cast_fp16 = transpose(perm = var_1081, x = x_87_cast_fp16)[name = string("transpose_118")];
+            tensor<fp16, [16, 192, 64]> key_layer_9_cast_fp16 = reshape(shape = var_1086, x = var_1082_cast_fp16)[name = string("key_layer_9_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376581824)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378679040)))];
-            tensor<fp16, [1, 256, 1024]> linear_34_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_34_cast_fp16")];
-            tensor<int32, [4]> var_1093 = const()[name = string("op_1093"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_91_cast_fp16 = reshape(shape = var_1093, x = linear_34_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_34_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<int32, [4]> var_1093 = const()[name = string("op_1093"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_91_cast_fp16 = reshape(shape = var_1093, x = linear_34_cast_fp16)[name = string("x_91_cast_fp16")];
             tensor<int32, [4]> var_1095 = const()[name = string("op_1095"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1096_cast_fp16 = transpose(perm = var_1095, x = x_91_cast_fp16)[name = string("transpose_117")];
-            tensor<fp16, [16, 256, 64]> value_layer_9_cast_fp16 = reshape(shape = var_1100, x = var_1096_cast_fp16)[name = string("value_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1096_cast_fp16 = transpose(perm = var_1095, x = x_91_cast_fp16)[name = string("transpose_117")];
+            tensor<fp16, [16, 192, 64]> value_layer_9_cast_fp16 = reshape(shape = var_1100, x = var_1096_cast_fp16)[name = string("value_layer_9_cast_fp16")];
             tensor<int32, [3]> var_1110_perm_0 = const()[name = string("op_1110_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1112_y_0_to_fp16 = const()[name = string("_inversed_1112_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_1110_cast_fp16 = transpose(perm = var_1110_perm_0, x = key_layer_9_cast_fp16)[name = string("transpose_116")];
-            tensor<fp16, [16, 64, 256]> _inversed_1112_cast_fp16 = mul(x = var_1110_cast_fp16, y = _inversed_1112_y_0_to_fp16)[name = string("_inversed_1112_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_1110_cast_fp16 = transpose(perm = var_1110_perm_0, x = key_layer_9_cast_fp16)[name = string("transpose_116")];
+            tensor<fp16, [16, 64, 192]> _inversed_1112_cast_fp16 = mul(x = var_1110_cast_fp16, y = _inversed_1112_y_0_to_fp16)[name = string("_inversed_1112_cast_fp16")];
             bool attention_scores_25_transpose_x_0 = const()[name = string("attention_scores_25_transpose_x_0"), val = bool(false)];
             bool attention_scores_25_transpose_y_0 = const()[name = string("attention_scores_25_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_0, transpose_y = attention_scores_25_transpose_y_0, x = query_layer_27_cast_fp16, y = _inversed_1112_cast_fp16)[name = string("attention_scores_25_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_0, transpose_y = attention_scores_25_transpose_y_0, x = query_layer_27_cast_fp16, y = _inversed_1112_cast_fp16)[name = string("attention_scores_25_cast_fp16")];
             bool c2p_att_17_transpose_x_0 = const()[name = string("c2p_att_17_transpose_x_0"), val = bool(false)];
             bool c2p_att_17_transpose_y_0 = const()[name = string("c2p_att_17_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378681152)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_17_cast_fp16 = matmul(transpose_x = c2p_att_17_transpose_x_0, transpose_y = c2p_att_17_transpose_y_0, x = query_layer_27_cast_fp16, y = var_1168_to_fp16)[name = string("c2p_att_17_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_17_cast_fp16 = matmul(transpose_x = c2p_att_17_transpose_x_0, transpose_y = c2p_att_17_transpose_y_0, x = query_layer_27_cast_fp16, y = var_1168_to_fp16)[name = string("c2p_att_17_cast_fp16")];
             bool c2p_att_19_validate_indices_0 = const()[name = string("c2p_att_19_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_19_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_19_validate_indices_0, x = c2p_att_17_cast_fp16)[name = string("c2p_att_19_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_19_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_19_validate_indices_0, x = c2p_att_17_cast_fp16)[name = string("c2p_att_19_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1181_y_0_to_fp16 = const()[name = string("_inversed_1181_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_1181_cast_fp16 = mul(x = c2p_att_19_cast_fp16_cast_int16, y = _inversed_1181_y_0_to_fp16)[name = string("_inversed_1181_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_1181_cast_fp16 = mul(x = c2p_att_19_cast_fp16_cast_int16, y = _inversed_1181_y_0_to_fp16)[name = string("_inversed_1181_cast_fp16")];
             bool p2c_att_17_transpose_x_0 = const()[name = string("p2c_att_17_transpose_x_0"), val = bool(false)];
             bool p2c_att_17_transpose_y_0 = const()[name = string("p2c_att_17_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1196_to_fp16 = const()[name = string("op_1196_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379729792)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_17_cast_fp16 = matmul(transpose_x = p2c_att_17_transpose_x_0, transpose_y = p2c_att_17_transpose_y_0, x = key_layer_9_cast_fp16, y = var_1196_to_fp16)[name = string("p2c_att_17_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_17_cast_fp16 = matmul(transpose_x = p2c_att_17_transpose_x_0, transpose_y = p2c_att_17_transpose_y_0, x = key_layer_9_cast_fp16, y = var_1196_to_fp16)[name = string("p2c_att_17_cast_fp16")];
             bool var_1204_validate_indices_0 = const()[name = string("op_1204_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_1204_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1204_validate_indices_0, x = p2c_att_17_cast_fp16)[name = string("op_1204_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_1204_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1204_validate_indices_0, x = p2c_att_17_cast_fp16)[name = string("op_1204_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_19_perm_0 = const()[name = string("p2c_att_19_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1207_y_0_to_fp16 = const()[name = string("_inversed_1207_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_19_cast_fp16 = transpose(perm = p2c_att_19_perm_0, x = var_1204_cast_fp16_cast_int16)[name = string("transpose_115")];
-            tensor<fp16, [16, 256, 256]> _inversed_1207_cast_fp16 = mul(x = p2c_att_19_cast_fp16, y = _inversed_1207_y_0_to_fp16)[name = string("_inversed_1207_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_9_cast_fp16 = add(x = _inversed_1181_cast_fp16, y = _inversed_1207_cast_fp16)[name = string("rel_att_9_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_27_cast_fp16 = add(x = attention_scores_25_cast_fp16, y = rel_att_9_cast_fp16)[name = string("attention_scores_27_cast_fp16")];
-            tensor<int32, [4]> var_1212 = const()[name = string("op_1212"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_29_cast_fp16 = reshape(shape = var_1212, x = attention_scores_27_cast_fp16)[name = string("attention_scores_29_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_105_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_29_cast_fp16, cond = var_339)[name = string("input_105_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_107_cast_fp16 = softmax(axis = var_56, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
-            tensor<int32, [3]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_1222_cast_fp16 = reshape(shape = var_1221, x = input_107_cast_fp16)[name = string("op_1222_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_19_cast_fp16 = transpose(perm = p2c_att_19_perm_0, x = var_1204_cast_fp16_cast_int16)[name = string("transpose_115")];
+            tensor<fp16, [16, 192, 192]> _inversed_1207_cast_fp16 = mul(x = p2c_att_19_cast_fp16, y = _inversed_1207_y_0_to_fp16)[name = string("_inversed_1207_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_9_cast_fp16 = add(x = _inversed_1181_cast_fp16, y = _inversed_1207_cast_fp16)[name = string("rel_att_9_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_27_cast_fp16 = add(x = attention_scores_25_cast_fp16, y = rel_att_9_cast_fp16)[name = string("attention_scores_27_cast_fp16")];
+            tensor<int32, [4]> var_1212 = const()[name = string("op_1212"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_29_cast_fp16 = reshape(shape = var_1212, x = attention_scores_27_cast_fp16)[name = string("attention_scores_29_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_105_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_29_cast_fp16, cond = var_339)[name = string("input_105_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_107_cast_fp16 = softmax(axis = var_56, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
+            tensor<int32, [3]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_1222_cast_fp16 = reshape(shape = var_1221, x = input_107_cast_fp16)[name = string("op_1222_cast_fp16")];
             bool context_layer_17_transpose_x_0 = const()[name = string("context_layer_17_transpose_x_0"), val = bool(false)];
             bool context_layer_17_transpose_y_0 = const()[name = string("context_layer_17_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = var_1222_cast_fp16, y = value_layer_9_cast_fp16)[name = string("context_layer_17_cast_fp16")];
-            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1227_cast_fp16 = reshape(shape = var_1226, x = context_layer_17_cast_fp16)[name = string("op_1227_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = var_1222_cast_fp16, y = value_layer_9_cast_fp16)[name = string("context_layer_17_cast_fp16")];
+            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1227_cast_fp16 = reshape(shape = var_1226, x = context_layer_17_cast_fp16)[name = string("op_1227_cast_fp16")];
             tensor<int32, [4]> var_1228 = const()[name = string("op_1228"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_1229_cast_fp16 = transpose(perm = var_1228, x = var_1227_cast_fp16)[name = string("transpose_114")];
-            tensor<fp16, [1, 256, 1024]> input_109_cast_fp16 = reshape(shape = var_1233, x = var_1229_cast_fp16)[name = string("input_109_cast_fp16")];
+            tensor<int32, [3]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_1229_cast_fp16 = transpose(perm = var_1228, x = var_1227_cast_fp16)[name = string("transpose_114")];
+            tensor<fp16, [1, 192, 1024]> input_109_cast_fp16 = reshape(shape = var_1233, x = var_1229_cast_fp16)[name = string("input_109_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380778432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382875648)))];
-            tensor<fp16, [1, 256, 1024]> linear_37_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_37_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_113_cast_fp16 = add(x = linear_37_cast_fp16, y = input_101_cast_fp16)[name = string("input_113_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_37_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_113_cast_fp16 = add(x = linear_37_cast_fp16, y = input_101_cast_fp16)[name = string("input_113_cast_fp16")];
             tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382877760)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382879872)))];
-            tensor<fp16, [1, 256, 1024]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_4_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382881984)))];
             tensor<fp16, [4096]> encoder_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391270656)))];
-            tensor<fp16, [1, 256, 4096]> linear_38_cast_fp16 = linear(bias = encoder_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_38_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_38_cast_fp16 = linear(bias = encoder_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_38_cast_fp16")];
             string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = string("input_119_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = string("input_119_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_4_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391278912)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399667584)))];
-            tensor<fp16, [1, 256, 1024]> linear_39_cast_fp16 = linear(bias = encoder_encoder_layer_4_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_output_dense_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_39_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_123_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = string("input_123_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_39_cast_fp16 = linear(bias = encoder_encoder_layer_4_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_output_dense_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_123_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = string("input_123_cast_fp16")];
             tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399669696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399671808)))];
-            tensor<fp16, [1, 256, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399673920)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401771136)))];
-            tensor<fp16, [1, 256, 1024]> linear_40_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_40_cast_fp16")];
-            tensor<int32, [4]> var_1284 = const()[name = string("op_1284"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_103_cast_fp16 = reshape(shape = var_1284, x = linear_40_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_40_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1284 = const()[name = string("op_1284"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_103_cast_fp16 = reshape(shape = var_1284, x = linear_40_cast_fp16)[name = string("x_103_cast_fp16")];
             tensor<int32, [4]> var_1286 = const()[name = string("op_1286"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1291 = const()[name = string("op_1291"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1287_cast_fp16 = transpose(perm = var_1286, x = x_103_cast_fp16)[name = string("transpose_113")];
-            tensor<fp16, [16, 256, 64]> query_layer_33_cast_fp16 = reshape(shape = var_1291, x = var_1287_cast_fp16)[name = string("query_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_1291 = const()[name = string("op_1291"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1287_cast_fp16 = transpose(perm = var_1286, x = x_103_cast_fp16)[name = string("transpose_113")];
+            tensor<fp16, [16, 192, 64]> query_layer_33_cast_fp16 = reshape(shape = var_1291, x = var_1287_cast_fp16)[name = string("query_layer_33_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401773248)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403870464)))];
-            tensor<fp16, [1, 256, 1024]> linear_41_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_41_cast_fp16")];
-            tensor<int32, [4]> var_1298 = const()[name = string("op_1298"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_107_cast_fp16 = reshape(shape = var_1298, x = linear_41_cast_fp16)[name = string("x_107_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_41_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1298 = const()[name = string("op_1298"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_107_cast_fp16 = reshape(shape = var_1298, x = linear_41_cast_fp16)[name = string("x_107_cast_fp16")];
             tensor<int32, [4]> var_1300 = const()[name = string("op_1300"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1305 = const()[name = string("op_1305"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1301_cast_fp16 = transpose(perm = var_1300, x = x_107_cast_fp16)[name = string("transpose_112")];
-            tensor<fp16, [16, 256, 64]> key_layer_11_cast_fp16 = reshape(shape = var_1305, x = var_1301_cast_fp16)[name = string("key_layer_11_cast_fp16")];
+            tensor<int32, [3]> var_1305 = const()[name = string("op_1305"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1301_cast_fp16 = transpose(perm = var_1300, x = x_107_cast_fp16)[name = string("transpose_112")];
+            tensor<fp16, [16, 192, 64]> key_layer_11_cast_fp16 = reshape(shape = var_1305, x = var_1301_cast_fp16)[name = string("key_layer_11_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403872576)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405969792)))];
-            tensor<fp16, [1, 256, 1024]> linear_42_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_42_cast_fp16")];
-            tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_111_cast_fp16 = reshape(shape = var_1312, x = linear_42_cast_fp16)[name = string("x_111_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_42_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_42_cast_fp16")];
+            tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_111_cast_fp16 = reshape(shape = var_1312, x = linear_42_cast_fp16)[name = string("x_111_cast_fp16")];
             tensor<int32, [4]> var_1314 = const()[name = string("op_1314"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1319 = const()[name = string("op_1319"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1315_cast_fp16 = transpose(perm = var_1314, x = x_111_cast_fp16)[name = string("transpose_111")];
-            tensor<fp16, [16, 256, 64]> value_layer_11_cast_fp16 = reshape(shape = var_1319, x = var_1315_cast_fp16)[name = string("value_layer_11_cast_fp16")];
+            tensor<int32, [3]> var_1319 = const()[name = string("op_1319"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1315_cast_fp16 = transpose(perm = var_1314, x = x_111_cast_fp16)[name = string("transpose_111")];
+            tensor<fp16, [16, 192, 64]> value_layer_11_cast_fp16 = reshape(shape = var_1319, x = var_1315_cast_fp16)[name = string("value_layer_11_cast_fp16")];
             tensor<int32, [3]> var_1329_perm_0 = const()[name = string("op_1329_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1331_y_0_to_fp16 = const()[name = string("_inversed_1331_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_1329_cast_fp16 = transpose(perm = var_1329_perm_0, x = key_layer_11_cast_fp16)[name = string("transpose_110")];
-            tensor<fp16, [16, 64, 256]> _inversed_1331_cast_fp16 = mul(x = var_1329_cast_fp16, y = _inversed_1331_y_0_to_fp16)[name = string("_inversed_1331_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_1329_cast_fp16 = transpose(perm = var_1329_perm_0, x = key_layer_11_cast_fp16)[name = string("transpose_110")];
+            tensor<fp16, [16, 64, 192]> _inversed_1331_cast_fp16 = mul(x = var_1329_cast_fp16, y = _inversed_1331_y_0_to_fp16)[name = string("_inversed_1331_cast_fp16")];
             bool attention_scores_31_transpose_x_0 = const()[name = string("attention_scores_31_transpose_x_0"), val = bool(false)];
             bool attention_scores_31_transpose_y_0 = const()[name = string("attention_scores_31_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_31_cast_fp16 = matmul(transpose_x = attention_scores_31_transpose_x_0, transpose_y = attention_scores_31_transpose_y_0, x = query_layer_33_cast_fp16, y = _inversed_1331_cast_fp16)[name = string("attention_scores_31_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_31_cast_fp16 = matmul(transpose_x = attention_scores_31_transpose_x_0, transpose_y = attention_scores_31_transpose_y_0, x = query_layer_33_cast_fp16, y = _inversed_1331_cast_fp16)[name = string("attention_scores_31_cast_fp16")];
             bool c2p_att_21_transpose_x_0 = const()[name = string("c2p_att_21_transpose_x_0"), val = bool(false)];
             bool c2p_att_21_transpose_y_0 = const()[name = string("c2p_att_21_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1387_to_fp16 = const()[name = string("op_1387_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405971904)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_21_cast_fp16 = matmul(transpose_x = c2p_att_21_transpose_x_0, transpose_y = c2p_att_21_transpose_y_0, x = query_layer_33_cast_fp16, y = var_1387_to_fp16)[name = string("c2p_att_21_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_21_cast_fp16 = matmul(transpose_x = c2p_att_21_transpose_x_0, transpose_y = c2p_att_21_transpose_y_0, x = query_layer_33_cast_fp16, y = var_1387_to_fp16)[name = string("c2p_att_21_cast_fp16")];
             bool c2p_att_23_validate_indices_0 = const()[name = string("c2p_att_23_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_23_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_23_validate_indices_0, x = c2p_att_21_cast_fp16)[name = string("c2p_att_23_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_23_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_23_validate_indices_0, x = c2p_att_21_cast_fp16)[name = string("c2p_att_23_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1400_y_0_to_fp16 = const()[name = string("_inversed_1400_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_1400_cast_fp16 = mul(x = c2p_att_23_cast_fp16_cast_int16, y = _inversed_1400_y_0_to_fp16)[name = string("_inversed_1400_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_1400_cast_fp16 = mul(x = c2p_att_23_cast_fp16_cast_int16, y = _inversed_1400_y_0_to_fp16)[name = string("_inversed_1400_cast_fp16")];
             bool p2c_att_21_transpose_x_0 = const()[name = string("p2c_att_21_transpose_x_0"), val = bool(false)];
             bool p2c_att_21_transpose_y_0 = const()[name = string("p2c_att_21_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1415_to_fp16 = const()[name = string("op_1415_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020544)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_21_cast_fp16 = matmul(transpose_x = p2c_att_21_transpose_x_0, transpose_y = p2c_att_21_transpose_y_0, x = key_layer_11_cast_fp16, y = var_1415_to_fp16)[name = string("p2c_att_21_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_21_cast_fp16 = matmul(transpose_x = p2c_att_21_transpose_x_0, transpose_y = p2c_att_21_transpose_y_0, x = key_layer_11_cast_fp16, y = var_1415_to_fp16)[name = string("p2c_att_21_cast_fp16")];
             bool var_1423_validate_indices_0 = const()[name = string("op_1423_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_1423_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1423_validate_indices_0, x = p2c_att_21_cast_fp16)[name = string("op_1423_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_1423_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1423_validate_indices_0, x = p2c_att_21_cast_fp16)[name = string("op_1423_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_23_perm_0 = const()[name = string("p2c_att_23_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1426_y_0_to_fp16 = const()[name = string("_inversed_1426_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_23_cast_fp16 = transpose(perm = p2c_att_23_perm_0, x = var_1423_cast_fp16_cast_int16)[name = string("transpose_109")];
-            tensor<fp16, [16, 256, 256]> _inversed_1426_cast_fp16 = mul(x = p2c_att_23_cast_fp16, y = _inversed_1426_y_0_to_fp16)[name = string("_inversed_1426_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_11_cast_fp16 = add(x = _inversed_1400_cast_fp16, y = _inversed_1426_cast_fp16)[name = string("rel_att_11_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_33_cast_fp16 = add(x = attention_scores_31_cast_fp16, y = rel_att_11_cast_fp16)[name = string("attention_scores_33_cast_fp16")];
-            tensor<int32, [4]> var_1431 = const()[name = string("op_1431"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_35_cast_fp16 = reshape(shape = var_1431, x = attention_scores_33_cast_fp16)[name = string("attention_scores_35_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_129_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_35_cast_fp16, cond = var_339)[name = string("input_129_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_131_cast_fp16 = softmax(axis = var_56, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
-            tensor<int32, [3]> var_1440 = const()[name = string("op_1440"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_1441_cast_fp16 = reshape(shape = var_1440, x = input_131_cast_fp16)[name = string("op_1441_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_23_cast_fp16 = transpose(perm = p2c_att_23_perm_0, x = var_1423_cast_fp16_cast_int16)[name = string("transpose_109")];
+            tensor<fp16, [16, 192, 192]> _inversed_1426_cast_fp16 = mul(x = p2c_att_23_cast_fp16, y = _inversed_1426_y_0_to_fp16)[name = string("_inversed_1426_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_11_cast_fp16 = add(x = _inversed_1400_cast_fp16, y = _inversed_1426_cast_fp16)[name = string("rel_att_11_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_33_cast_fp16 = add(x = attention_scores_31_cast_fp16, y = rel_att_11_cast_fp16)[name = string("attention_scores_33_cast_fp16")];
+            tensor<int32, [4]> var_1431 = const()[name = string("op_1431"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_35_cast_fp16 = reshape(shape = var_1431, x = attention_scores_33_cast_fp16)[name = string("attention_scores_35_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_129_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_35_cast_fp16, cond = var_339)[name = string("input_129_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_131_cast_fp16 = softmax(axis = var_56, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<int32, [3]> var_1440 = const()[name = string("op_1440"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_1441_cast_fp16 = reshape(shape = var_1440, x = input_131_cast_fp16)[name = string("op_1441_cast_fp16")];
             bool context_layer_21_transpose_x_0 = const()[name = string("context_layer_21_transpose_x_0"), val = bool(false)];
             bool context_layer_21_transpose_y_0 = const()[name = string("context_layer_21_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = var_1441_cast_fp16, y = value_layer_11_cast_fp16)[name = string("context_layer_21_cast_fp16")];
-            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1446_cast_fp16 = reshape(shape = var_1445, x = context_layer_21_cast_fp16)[name = string("op_1446_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = var_1441_cast_fp16, y = value_layer_11_cast_fp16)[name = string("context_layer_21_cast_fp16")];
+            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1446_cast_fp16 = reshape(shape = var_1445, x = context_layer_21_cast_fp16)[name = string("op_1446_cast_fp16")];
             tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_1448_cast_fp16 = transpose(perm = var_1447, x = var_1446_cast_fp16)[name = string("transpose_108")];
-            tensor<fp16, [1, 256, 1024]> input_133_cast_fp16 = reshape(shape = var_1452, x = var_1448_cast_fp16)[name = string("input_133_cast_fp16")];
+            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_1448_cast_fp16 = transpose(perm = var_1447, x = var_1446_cast_fp16)[name = string("transpose_108")];
+            tensor<fp16, [1, 192, 1024]> input_133_cast_fp16 = reshape(shape = var_1452, x = var_1448_cast_fp16)[name = string("input_133_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408069184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410166400)))];
-            tensor<fp16, [1, 256, 1024]> linear_45_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_45_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_125_cast_fp16)[name = string("input_137_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_45_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_45_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_125_cast_fp16)[name = string("input_137_cast_fp16")];
             tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410168512)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410170624)))];
-            tensor<fp16, [1, 256, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_5_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410172736)))];
             tensor<fp16, [4096]> encoder_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418561408)))];
-            tensor<fp16, [1, 256, 4096]> linear_46_cast_fp16 = linear(bias = encoder_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_46_cast_fp16 = linear(bias = encoder_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_46_cast_fp16")];
             string input_143_mode_0 = const()[name = string("input_143_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = string("input_143_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = string("input_143_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_5_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418569664)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426958336)))];
-            tensor<fp16, [1, 256, 1024]> linear_47_cast_fp16 = linear(bias = encoder_encoder_layer_5_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_output_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_47_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_139_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_47_cast_fp16 = linear(bias = encoder_encoder_layer_5_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_output_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_47_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_139_cast_fp16)[name = string("input_147_cast_fp16")];
             tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426960448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426962560)))];
-            tensor<fp16, [1, 256, 1024]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426964672)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429061888)))];
-            tensor<fp16, [1, 256, 1024]> linear_48_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_48_cast_fp16")];
-            tensor<int32, [4]> var_1503 = const()[name = string("op_1503"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_123_cast_fp16 = reshape(shape = var_1503, x = linear_48_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_48_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1503 = const()[name = string("op_1503"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_123_cast_fp16 = reshape(shape = var_1503, x = linear_48_cast_fp16)[name = string("x_123_cast_fp16")];
             tensor<int32, [4]> var_1505 = const()[name = string("op_1505"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1506_cast_fp16 = transpose(perm = var_1505, x = x_123_cast_fp16)[name = string("transpose_107")];
-            tensor<fp16, [16, 256, 64]> query_layer_39_cast_fp16 = reshape(shape = var_1510, x = var_1506_cast_fp16)[name = string("query_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1506_cast_fp16 = transpose(perm = var_1505, x = x_123_cast_fp16)[name = string("transpose_107")];
+            tensor<fp16, [16, 192, 64]> query_layer_39_cast_fp16 = reshape(shape = var_1510, x = var_1506_cast_fp16)[name = string("query_layer_39_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429064000)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431161216)))];
-            tensor<fp16, [1, 256, 1024]> linear_49_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_49_cast_fp16")];
-            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_127_cast_fp16 = reshape(shape = var_1517, x = linear_49_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_49_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_127_cast_fp16 = reshape(shape = var_1517, x = linear_49_cast_fp16)[name = string("x_127_cast_fp16")];
             tensor<int32, [4]> var_1519 = const()[name = string("op_1519"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1520_cast_fp16 = transpose(perm = var_1519, x = x_127_cast_fp16)[name = string("transpose_106")];
-            tensor<fp16, [16, 256, 64]> key_layer_13_cast_fp16 = reshape(shape = var_1524, x = var_1520_cast_fp16)[name = string("key_layer_13_cast_fp16")];
+            tensor<int32, [3]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1520_cast_fp16 = transpose(perm = var_1519, x = x_127_cast_fp16)[name = string("transpose_106")];
+            tensor<fp16, [16, 192, 64]> key_layer_13_cast_fp16 = reshape(shape = var_1524, x = var_1520_cast_fp16)[name = string("key_layer_13_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431163328)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433260544)))];
-            tensor<fp16, [1, 256, 1024]> linear_50_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_50_cast_fp16")];
-            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_131_cast_fp16 = reshape(shape = var_1531, x = linear_50_cast_fp16)[name = string("x_131_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_50_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_131_cast_fp16 = reshape(shape = var_1531, x = linear_50_cast_fp16)[name = string("x_131_cast_fp16")];
             tensor<int32, [4]> var_1533 = const()[name = string("op_1533"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1538 = const()[name = string("op_1538"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1534_cast_fp16 = transpose(perm = var_1533, x = x_131_cast_fp16)[name = string("transpose_105")];
-            tensor<fp16, [16, 256, 64]> value_layer_13_cast_fp16 = reshape(shape = var_1538, x = var_1534_cast_fp16)[name = string("value_layer_13_cast_fp16")];
+            tensor<int32, [3]> var_1538 = const()[name = string("op_1538"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1534_cast_fp16 = transpose(perm = var_1533, x = x_131_cast_fp16)[name = string("transpose_105")];
+            tensor<fp16, [16, 192, 64]> value_layer_13_cast_fp16 = reshape(shape = var_1538, x = var_1534_cast_fp16)[name = string("value_layer_13_cast_fp16")];
             tensor<int32, [3]> var_1548_perm_0 = const()[name = string("op_1548_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1550_y_0_to_fp16 = const()[name = string("_inversed_1550_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_1548_cast_fp16 = transpose(perm = var_1548_perm_0, x = key_layer_13_cast_fp16)[name = string("transpose_104")];
-            tensor<fp16, [16, 64, 256]> _inversed_1550_cast_fp16 = mul(x = var_1548_cast_fp16, y = _inversed_1550_y_0_to_fp16)[name = string("_inversed_1550_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_1548_cast_fp16 = transpose(perm = var_1548_perm_0, x = key_layer_13_cast_fp16)[name = string("transpose_104")];
+            tensor<fp16, [16, 64, 192]> _inversed_1550_cast_fp16 = mul(x = var_1548_cast_fp16, y = _inversed_1550_y_0_to_fp16)[name = string("_inversed_1550_cast_fp16")];
             bool attention_scores_37_transpose_x_0 = const()[name = string("attention_scores_37_transpose_x_0"), val = bool(false)];
             bool attention_scores_37_transpose_y_0 = const()[name = string("attention_scores_37_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_37_cast_fp16 = matmul(transpose_x = attention_scores_37_transpose_x_0, transpose_y = attention_scores_37_transpose_y_0, x = query_layer_39_cast_fp16, y = _inversed_1550_cast_fp16)[name = string("attention_scores_37_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_37_cast_fp16 = matmul(transpose_x = attention_scores_37_transpose_x_0, transpose_y = attention_scores_37_transpose_y_0, x = query_layer_39_cast_fp16, y = _inversed_1550_cast_fp16)[name = string("attention_scores_37_cast_fp16")];
             bool c2p_att_25_transpose_x_0 = const()[name = string("c2p_att_25_transpose_x_0"), val = bool(false)];
             bool c2p_att_25_transpose_y_0 = const()[name = string("c2p_att_25_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1606_to_fp16 = const()[name = string("op_1606_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433262656)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_25_cast_fp16 = matmul(transpose_x = c2p_att_25_transpose_x_0, transpose_y = c2p_att_25_transpose_y_0, x = query_layer_39_cast_fp16, y = var_1606_to_fp16)[name = string("c2p_att_25_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_25_cast_fp16 = matmul(transpose_x = c2p_att_25_transpose_x_0, transpose_y = c2p_att_25_transpose_y_0, x = query_layer_39_cast_fp16, y = var_1606_to_fp16)[name = string("c2p_att_25_cast_fp16")];
             bool c2p_att_27_validate_indices_0 = const()[name = string("c2p_att_27_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_27_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_27_validate_indices_0, x = c2p_att_25_cast_fp16)[name = string("c2p_att_27_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_27_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_27_validate_indices_0, x = c2p_att_25_cast_fp16)[name = string("c2p_att_27_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1619_y_0_to_fp16 = const()[name = string("_inversed_1619_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_1619_cast_fp16 = mul(x = c2p_att_27_cast_fp16_cast_int16, y = _inversed_1619_y_0_to_fp16)[name = string("_inversed_1619_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_1619_cast_fp16 = mul(x = c2p_att_27_cast_fp16_cast_int16, y = _inversed_1619_y_0_to_fp16)[name = string("_inversed_1619_cast_fp16")];
             bool p2c_att_25_transpose_x_0 = const()[name = string("p2c_att_25_transpose_x_0"), val = bool(false)];
             bool p2c_att_25_transpose_y_0 = const()[name = string("p2c_att_25_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1634_to_fp16 = const()[name = string("op_1634_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434311296)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_25_cast_fp16 = matmul(transpose_x = p2c_att_25_transpose_x_0, transpose_y = p2c_att_25_transpose_y_0, x = key_layer_13_cast_fp16, y = var_1634_to_fp16)[name = string("p2c_att_25_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_25_cast_fp16 = matmul(transpose_x = p2c_att_25_transpose_x_0, transpose_y = p2c_att_25_transpose_y_0, x = key_layer_13_cast_fp16, y = var_1634_to_fp16)[name = string("p2c_att_25_cast_fp16")];
             bool var_1642_validate_indices_0 = const()[name = string("op_1642_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_1642_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1642_validate_indices_0, x = p2c_att_25_cast_fp16)[name = string("op_1642_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_1642_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1642_validate_indices_0, x = p2c_att_25_cast_fp16)[name = string("op_1642_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_27_perm_0 = const()[name = string("p2c_att_27_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1645_y_0_to_fp16 = const()[name = string("_inversed_1645_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_27_cast_fp16 = transpose(perm = p2c_att_27_perm_0, x = var_1642_cast_fp16_cast_int16)[name = string("transpose_103")];
-            tensor<fp16, [16, 256, 256]> _inversed_1645_cast_fp16 = mul(x = p2c_att_27_cast_fp16, y = _inversed_1645_y_0_to_fp16)[name = string("_inversed_1645_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_13_cast_fp16 = add(x = _inversed_1619_cast_fp16, y = _inversed_1645_cast_fp16)[name = string("rel_att_13_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_39_cast_fp16 = add(x = attention_scores_37_cast_fp16, y = rel_att_13_cast_fp16)[name = string("attention_scores_39_cast_fp16")];
-            tensor<int32, [4]> var_1650 = const()[name = string("op_1650"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_41_cast_fp16 = reshape(shape = var_1650, x = attention_scores_39_cast_fp16)[name = string("attention_scores_41_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_153_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_41_cast_fp16, cond = var_339)[name = string("input_153_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_155_cast_fp16 = softmax(axis = var_56, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
-            tensor<int32, [3]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_1660_cast_fp16 = reshape(shape = var_1659, x = input_155_cast_fp16)[name = string("op_1660_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_27_cast_fp16 = transpose(perm = p2c_att_27_perm_0, x = var_1642_cast_fp16_cast_int16)[name = string("transpose_103")];
+            tensor<fp16, [16, 192, 192]> _inversed_1645_cast_fp16 = mul(x = p2c_att_27_cast_fp16, y = _inversed_1645_y_0_to_fp16)[name = string("_inversed_1645_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_13_cast_fp16 = add(x = _inversed_1619_cast_fp16, y = _inversed_1645_cast_fp16)[name = string("rel_att_13_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_39_cast_fp16 = add(x = attention_scores_37_cast_fp16, y = rel_att_13_cast_fp16)[name = string("attention_scores_39_cast_fp16")];
+            tensor<int32, [4]> var_1650 = const()[name = string("op_1650"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_41_cast_fp16 = reshape(shape = var_1650, x = attention_scores_39_cast_fp16)[name = string("attention_scores_41_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_153_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_41_cast_fp16, cond = var_339)[name = string("input_153_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_155_cast_fp16 = softmax(axis = var_56, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            tensor<int32, [3]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_1660_cast_fp16 = reshape(shape = var_1659, x = input_155_cast_fp16)[name = string("op_1660_cast_fp16")];
             bool context_layer_25_transpose_x_0 = const()[name = string("context_layer_25_transpose_x_0"), val = bool(false)];
             bool context_layer_25_transpose_y_0 = const()[name = string("context_layer_25_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = var_1660_cast_fp16, y = value_layer_13_cast_fp16)[name = string("context_layer_25_cast_fp16")];
-            tensor<int32, [4]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1665_cast_fp16 = reshape(shape = var_1664, x = context_layer_25_cast_fp16)[name = string("op_1665_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = var_1660_cast_fp16, y = value_layer_13_cast_fp16)[name = string("context_layer_25_cast_fp16")];
+            tensor<int32, [4]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1665_cast_fp16 = reshape(shape = var_1664, x = context_layer_25_cast_fp16)[name = string("op_1665_cast_fp16")];
             tensor<int32, [4]> var_1666 = const()[name = string("op_1666"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1671 = const()[name = string("op_1671"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1666, x = var_1665_cast_fp16)[name = string("transpose_102")];
-            tensor<fp16, [1, 256, 1024]> input_157_cast_fp16 = reshape(shape = var_1671, x = var_1667_cast_fp16)[name = string("input_157_cast_fp16")];
+            tensor<int32, [3]> var_1671 = const()[name = string("op_1671"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1666, x = var_1665_cast_fp16)[name = string("transpose_102")];
+            tensor<fp16, [1, 192, 1024]> input_157_cast_fp16 = reshape(shape = var_1671, x = var_1667_cast_fp16)[name = string("input_157_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435359936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437457152)))];
-            tensor<fp16, [1, 256, 1024]> linear_53_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_53_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_161_cast_fp16 = add(x = linear_53_cast_fp16, y = input_149_cast_fp16)[name = string("input_161_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_53_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_161_cast_fp16 = add(x = linear_53_cast_fp16, y = input_149_cast_fp16)[name = string("input_161_cast_fp16")];
             tensor<int32, [1]> input_163_axes_0 = const()[name = string("input_163_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437459264)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437461376)))];
-            tensor<fp16, [1, 256, 1024]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_6_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437463488)))];
             tensor<fp16, [4096]> encoder_encoder_layer_6_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445852160)))];
-            tensor<fp16, [1, 256, 4096]> linear_54_cast_fp16 = linear(bias = encoder_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_54_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_54_cast_fp16 = linear(bias = encoder_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_54_cast_fp16")];
             string input_167_mode_0 = const()[name = string("input_167_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = string("input_167_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_6_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445860416)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454249088)))];
-            tensor<fp16, [1, 256, 1024]> linear_55_cast_fp16 = linear(bias = encoder_encoder_layer_6_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_output_dense_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_55_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_171_cast_fp16 = add(x = linear_55_cast_fp16, y = input_163_cast_fp16)[name = string("input_171_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_55_cast_fp16 = linear(bias = encoder_encoder_layer_6_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_output_dense_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_171_cast_fp16 = add(x = linear_55_cast_fp16, y = input_163_cast_fp16)[name = string("input_171_cast_fp16")];
             tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454251200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454253312)))];
-            tensor<fp16, [1, 256, 1024]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454255424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456352640)))];
-            tensor<fp16, [1, 256, 1024]> linear_56_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_56_cast_fp16")];
-            tensor<int32, [4]> var_1722 = const()[name = string("op_1722"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_143_cast_fp16 = reshape(shape = var_1722, x = linear_56_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_56_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_56_cast_fp16")];
+            tensor<int32, [4]> var_1722 = const()[name = string("op_1722"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_143_cast_fp16 = reshape(shape = var_1722, x = linear_56_cast_fp16)[name = string("x_143_cast_fp16")];
             tensor<int32, [4]> var_1724 = const()[name = string("op_1724"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1729 = const()[name = string("op_1729"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1725_cast_fp16 = transpose(perm = var_1724, x = x_143_cast_fp16)[name = string("transpose_101")];
-            tensor<fp16, [16, 256, 64]> query_layer_45_cast_fp16 = reshape(shape = var_1729, x = var_1725_cast_fp16)[name = string("query_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_1729 = const()[name = string("op_1729"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1725_cast_fp16 = transpose(perm = var_1724, x = x_143_cast_fp16)[name = string("transpose_101")];
+            tensor<fp16, [16, 192, 64]> query_layer_45_cast_fp16 = reshape(shape = var_1729, x = var_1725_cast_fp16)[name = string("query_layer_45_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456354752)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458451968)))];
-            tensor<fp16, [1, 256, 1024]> linear_57_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_57_cast_fp16")];
-            tensor<int32, [4]> var_1736 = const()[name = string("op_1736"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_147_cast_fp16 = reshape(shape = var_1736, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_57_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1736 = const()[name = string("op_1736"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_147_cast_fp16 = reshape(shape = var_1736, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
             tensor<int32, [4]> var_1738 = const()[name = string("op_1738"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1739_cast_fp16 = transpose(perm = var_1738, x = x_147_cast_fp16)[name = string("transpose_100")];
-            tensor<fp16, [16, 256, 64]> key_layer_15_cast_fp16 = reshape(shape = var_1743, x = var_1739_cast_fp16)[name = string("key_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1739_cast_fp16 = transpose(perm = var_1738, x = x_147_cast_fp16)[name = string("transpose_100")];
+            tensor<fp16, [16, 192, 64]> key_layer_15_cast_fp16 = reshape(shape = var_1743, x = var_1739_cast_fp16)[name = string("key_layer_15_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458454080)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460551296)))];
-            tensor<fp16, [1, 256, 1024]> linear_58_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_58_cast_fp16")];
-            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_151_cast_fp16 = reshape(shape = var_1750, x = linear_58_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_58_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_151_cast_fp16 = reshape(shape = var_1750, x = linear_58_cast_fp16)[name = string("x_151_cast_fp16")];
             tensor<int32, [4]> var_1752 = const()[name = string("op_1752"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1757 = const()[name = string("op_1757"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1753_cast_fp16 = transpose(perm = var_1752, x = x_151_cast_fp16)[name = string("transpose_99")];
-            tensor<fp16, [16, 256, 64]> value_layer_15_cast_fp16 = reshape(shape = var_1757, x = var_1753_cast_fp16)[name = string("value_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_1757 = const()[name = string("op_1757"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1753_cast_fp16 = transpose(perm = var_1752, x = x_151_cast_fp16)[name = string("transpose_99")];
+            tensor<fp16, [16, 192, 64]> value_layer_15_cast_fp16 = reshape(shape = var_1757, x = var_1753_cast_fp16)[name = string("value_layer_15_cast_fp16")];
             tensor<int32, [3]> var_1767_perm_0 = const()[name = string("op_1767_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1769_y_0_to_fp16 = const()[name = string("_inversed_1769_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = key_layer_15_cast_fp16)[name = string("transpose_98")];
-            tensor<fp16, [16, 64, 256]> _inversed_1769_cast_fp16 = mul(x = var_1767_cast_fp16, y = _inversed_1769_y_0_to_fp16)[name = string("_inversed_1769_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = key_layer_15_cast_fp16)[name = string("transpose_98")];
+            tensor<fp16, [16, 64, 192]> _inversed_1769_cast_fp16 = mul(x = var_1767_cast_fp16, y = _inversed_1769_y_0_to_fp16)[name = string("_inversed_1769_cast_fp16")];
             bool attention_scores_43_transpose_x_0 = const()[name = string("attention_scores_43_transpose_x_0"), val = bool(false)];
             bool attention_scores_43_transpose_y_0 = const()[name = string("attention_scores_43_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_43_cast_fp16 = matmul(transpose_x = attention_scores_43_transpose_x_0, transpose_y = attention_scores_43_transpose_y_0, x = query_layer_45_cast_fp16, y = _inversed_1769_cast_fp16)[name = string("attention_scores_43_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_43_cast_fp16 = matmul(transpose_x = attention_scores_43_transpose_x_0, transpose_y = attention_scores_43_transpose_y_0, x = query_layer_45_cast_fp16, y = _inversed_1769_cast_fp16)[name = string("attention_scores_43_cast_fp16")];
             bool c2p_att_29_transpose_x_0 = const()[name = string("c2p_att_29_transpose_x_0"), val = bool(false)];
             bool c2p_att_29_transpose_y_0 = const()[name = string("c2p_att_29_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1825_to_fp16 = const()[name = string("op_1825_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460553408)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_29_cast_fp16 = matmul(transpose_x = c2p_att_29_transpose_x_0, transpose_y = c2p_att_29_transpose_y_0, x = query_layer_45_cast_fp16, y = var_1825_to_fp16)[name = string("c2p_att_29_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_29_cast_fp16 = matmul(transpose_x = c2p_att_29_transpose_x_0, transpose_y = c2p_att_29_transpose_y_0, x = query_layer_45_cast_fp16, y = var_1825_to_fp16)[name = string("c2p_att_29_cast_fp16")];
             bool c2p_att_31_validate_indices_0 = const()[name = string("c2p_att_31_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_31_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_31_validate_indices_0, x = c2p_att_29_cast_fp16)[name = string("c2p_att_31_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_31_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_31_validate_indices_0, x = c2p_att_29_cast_fp16)[name = string("c2p_att_31_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1838_y_0_to_fp16 = const()[name = string("_inversed_1838_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_1838_cast_fp16 = mul(x = c2p_att_31_cast_fp16_cast_int16, y = _inversed_1838_y_0_to_fp16)[name = string("_inversed_1838_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_1838_cast_fp16 = mul(x = c2p_att_31_cast_fp16_cast_int16, y = _inversed_1838_y_0_to_fp16)[name = string("_inversed_1838_cast_fp16")];
             bool p2c_att_29_transpose_x_0 = const()[name = string("p2c_att_29_transpose_x_0"), val = bool(false)];
             bool p2c_att_29_transpose_y_0 = const()[name = string("p2c_att_29_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1853_to_fp16 = const()[name = string("op_1853_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461602048)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_29_cast_fp16 = matmul(transpose_x = p2c_att_29_transpose_x_0, transpose_y = p2c_att_29_transpose_y_0, x = key_layer_15_cast_fp16, y = var_1853_to_fp16)[name = string("p2c_att_29_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_29_cast_fp16 = matmul(transpose_x = p2c_att_29_transpose_x_0, transpose_y = p2c_att_29_transpose_y_0, x = key_layer_15_cast_fp16, y = var_1853_to_fp16)[name = string("p2c_att_29_cast_fp16")];
             bool var_1861_validate_indices_0 = const()[name = string("op_1861_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_1861_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1861_validate_indices_0, x = p2c_att_29_cast_fp16)[name = string("op_1861_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_1861_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1861_validate_indices_0, x = p2c_att_29_cast_fp16)[name = string("op_1861_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_31_perm_0 = const()[name = string("p2c_att_31_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1864_y_0_to_fp16 = const()[name = string("_inversed_1864_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_31_cast_fp16 = transpose(perm = p2c_att_31_perm_0, x = var_1861_cast_fp16_cast_int16)[name = string("transpose_97")];
-            tensor<fp16, [16, 256, 256]> _inversed_1864_cast_fp16 = mul(x = p2c_att_31_cast_fp16, y = _inversed_1864_y_0_to_fp16)[name = string("_inversed_1864_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_15_cast_fp16 = add(x = _inversed_1838_cast_fp16, y = _inversed_1864_cast_fp16)[name = string("rel_att_15_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_45_cast_fp16 = add(x = attention_scores_43_cast_fp16, y = rel_att_15_cast_fp16)[name = string("attention_scores_45_cast_fp16")];
-            tensor<int32, [4]> var_1869 = const()[name = string("op_1869"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_47_cast_fp16 = reshape(shape = var_1869, x = attention_scores_45_cast_fp16)[name = string("attention_scores_47_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_177_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_47_cast_fp16, cond = var_339)[name = string("input_177_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_179_cast_fp16 = softmax(axis = var_56, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
-            tensor<int32, [3]> var_1878 = const()[name = string("op_1878"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_1879_cast_fp16 = reshape(shape = var_1878, x = input_179_cast_fp16)[name = string("op_1879_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_31_cast_fp16 = transpose(perm = p2c_att_31_perm_0, x = var_1861_cast_fp16_cast_int16)[name = string("transpose_97")];
+            tensor<fp16, [16, 192, 192]> _inversed_1864_cast_fp16 = mul(x = p2c_att_31_cast_fp16, y = _inversed_1864_y_0_to_fp16)[name = string("_inversed_1864_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_15_cast_fp16 = add(x = _inversed_1838_cast_fp16, y = _inversed_1864_cast_fp16)[name = string("rel_att_15_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_45_cast_fp16 = add(x = attention_scores_43_cast_fp16, y = rel_att_15_cast_fp16)[name = string("attention_scores_45_cast_fp16")];
+            tensor<int32, [4]> var_1869 = const()[name = string("op_1869"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_47_cast_fp16 = reshape(shape = var_1869, x = attention_scores_45_cast_fp16)[name = string("attention_scores_47_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_177_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_47_cast_fp16, cond = var_339)[name = string("input_177_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_179_cast_fp16 = softmax(axis = var_56, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [3]> var_1878 = const()[name = string("op_1878"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_1879_cast_fp16 = reshape(shape = var_1878, x = input_179_cast_fp16)[name = string("op_1879_cast_fp16")];
             bool context_layer_29_transpose_x_0 = const()[name = string("context_layer_29_transpose_x_0"), val = bool(false)];
             bool context_layer_29_transpose_y_0 = const()[name = string("context_layer_29_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = var_1879_cast_fp16, y = value_layer_15_cast_fp16)[name = string("context_layer_29_cast_fp16")];
-            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1884_cast_fp16 = reshape(shape = var_1883, x = context_layer_29_cast_fp16)[name = string("op_1884_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = var_1879_cast_fp16, y = value_layer_15_cast_fp16)[name = string("context_layer_29_cast_fp16")];
+            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1884_cast_fp16 = reshape(shape = var_1883, x = context_layer_29_cast_fp16)[name = string("op_1884_cast_fp16")];
             tensor<int32, [4]> var_1885 = const()[name = string("op_1885"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1890 = const()[name = string("op_1890"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_1886_cast_fp16 = transpose(perm = var_1885, x = var_1884_cast_fp16)[name = string("transpose_96")];
-            tensor<fp16, [1, 256, 1024]> input_181_cast_fp16 = reshape(shape = var_1890, x = var_1886_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [3]> var_1890 = const()[name = string("op_1890"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_1886_cast_fp16 = transpose(perm = var_1885, x = var_1884_cast_fp16)[name = string("transpose_96")];
+            tensor<fp16, [1, 192, 1024]> input_181_cast_fp16 = reshape(shape = var_1890, x = var_1886_cast_fp16)[name = string("input_181_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462650688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464747904)))];
-            tensor<fp16, [1, 256, 1024]> linear_61_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_61_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_185_cast_fp16 = add(x = linear_61_cast_fp16, y = input_173_cast_fp16)[name = string("input_185_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_61_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_185_cast_fp16 = add(x = linear_61_cast_fp16, y = input_173_cast_fp16)[name = string("input_185_cast_fp16")];
             tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464750016)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464752128)))];
-            tensor<fp16, [1, 256, 1024]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_7_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464754240)))];
             tensor<fp16, [4096]> encoder_encoder_layer_7_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473142912)))];
-            tensor<fp16, [1, 256, 4096]> linear_62_cast_fp16 = linear(bias = encoder_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_62_cast_fp16 = linear(bias = encoder_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_62_cast_fp16")];
             string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = string("input_191_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = string("input_191_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_7_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473151168)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481539840)))];
-            tensor<fp16, [1, 256, 1024]> linear_63_cast_fp16 = linear(bias = encoder_encoder_layer_7_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_output_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_63_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_195_cast_fp16 = add(x = linear_63_cast_fp16, y = input_187_cast_fp16)[name = string("input_195_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_63_cast_fp16 = linear(bias = encoder_encoder_layer_7_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_output_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_63_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_195_cast_fp16 = add(x = linear_63_cast_fp16, y = input_187_cast_fp16)[name = string("input_195_cast_fp16")];
             tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481541952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481544064)))];
-            tensor<fp16, [1, 256, 1024]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481546176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483643392)))];
-            tensor<fp16, [1, 256, 1024]> linear_64_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_64_cast_fp16")];
-            tensor<int32, [4]> var_1941 = const()[name = string("op_1941"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_163_cast_fp16 = reshape(shape = var_1941, x = linear_64_cast_fp16)[name = string("x_163_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_64_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<int32, [4]> var_1941 = const()[name = string("op_1941"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_163_cast_fp16 = reshape(shape = var_1941, x = linear_64_cast_fp16)[name = string("x_163_cast_fp16")];
             tensor<int32, [4]> var_1943 = const()[name = string("op_1943"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1948 = const()[name = string("op_1948"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1944_cast_fp16 = transpose(perm = var_1943, x = x_163_cast_fp16)[name = string("transpose_95")];
-            tensor<fp16, [16, 256, 64]> query_layer_51_cast_fp16 = reshape(shape = var_1948, x = var_1944_cast_fp16)[name = string("query_layer_51_cast_fp16")];
+            tensor<int32, [3]> var_1948 = const()[name = string("op_1948"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1944_cast_fp16 = transpose(perm = var_1943, x = x_163_cast_fp16)[name = string("transpose_95")];
+            tensor<fp16, [16, 192, 64]> query_layer_51_cast_fp16 = reshape(shape = var_1948, x = var_1944_cast_fp16)[name = string("query_layer_51_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483645504)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485742720)))];
-            tensor<fp16, [1, 256, 1024]> linear_65_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_65_cast_fp16")];
-            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_167_cast_fp16 = reshape(shape = var_1955, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_65_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_65_cast_fp16")];
+            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_167_cast_fp16 = reshape(shape = var_1955, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
             tensor<int32, [4]> var_1957 = const()[name = string("op_1957"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1962 = const()[name = string("op_1962"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1958_cast_fp16 = transpose(perm = var_1957, x = x_167_cast_fp16)[name = string("transpose_94")];
-            tensor<fp16, [16, 256, 64]> key_layer_17_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = string("key_layer_17_cast_fp16")];
+            tensor<int32, [3]> var_1962 = const()[name = string("op_1962"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1958_cast_fp16 = transpose(perm = var_1957, x = x_167_cast_fp16)[name = string("transpose_94")];
+            tensor<fp16, [16, 192, 64]> key_layer_17_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = string("key_layer_17_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485744832)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487842048)))];
-            tensor<fp16, [1, 256, 1024]> linear_66_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_66_cast_fp16")];
-            tensor<int32, [4]> var_1969 = const()[name = string("op_1969"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_171_cast_fp16 = reshape(shape = var_1969, x = linear_66_cast_fp16)[name = string("x_171_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_66_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1969 = const()[name = string("op_1969"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_171_cast_fp16 = reshape(shape = var_1969, x = linear_66_cast_fp16)[name = string("x_171_cast_fp16")];
             tensor<int32, [4]> var_1971 = const()[name = string("op_1971"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_1972_cast_fp16 = transpose(perm = var_1971, x = x_171_cast_fp16)[name = string("transpose_93")];
-            tensor<fp16, [16, 256, 64]> value_layer_17_cast_fp16 = reshape(shape = var_1976, x = var_1972_cast_fp16)[name = string("value_layer_17_cast_fp16")];
+            tensor<int32, [3]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_1972_cast_fp16 = transpose(perm = var_1971, x = x_171_cast_fp16)[name = string("transpose_93")];
+            tensor<fp16, [16, 192, 64]> value_layer_17_cast_fp16 = reshape(shape = var_1976, x = var_1972_cast_fp16)[name = string("value_layer_17_cast_fp16")];
             tensor<int32, [3]> var_1986_perm_0 = const()[name = string("op_1986_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1988_y_0_to_fp16 = const()[name = string("_inversed_1988_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_1986_cast_fp16 = transpose(perm = var_1986_perm_0, x = key_layer_17_cast_fp16)[name = string("transpose_92")];
-            tensor<fp16, [16, 64, 256]> _inversed_1988_cast_fp16 = mul(x = var_1986_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = string("_inversed_1988_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_1986_cast_fp16 = transpose(perm = var_1986_perm_0, x = key_layer_17_cast_fp16)[name = string("transpose_92")];
+            tensor<fp16, [16, 64, 192]> _inversed_1988_cast_fp16 = mul(x = var_1986_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = string("_inversed_1988_cast_fp16")];
             bool attention_scores_49_transpose_x_0 = const()[name = string("attention_scores_49_transpose_x_0"), val = bool(false)];
             bool attention_scores_49_transpose_y_0 = const()[name = string("attention_scores_49_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_49_cast_fp16 = matmul(transpose_x = attention_scores_49_transpose_x_0, transpose_y = attention_scores_49_transpose_y_0, x = query_layer_51_cast_fp16, y = _inversed_1988_cast_fp16)[name = string("attention_scores_49_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_49_cast_fp16 = matmul(transpose_x = attention_scores_49_transpose_x_0, transpose_y = attention_scores_49_transpose_y_0, x = query_layer_51_cast_fp16, y = _inversed_1988_cast_fp16)[name = string("attention_scores_49_cast_fp16")];
             bool c2p_att_33_transpose_x_0 = const()[name = string("c2p_att_33_transpose_x_0"), val = bool(false)];
             bool c2p_att_33_transpose_y_0 = const()[name = string("c2p_att_33_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2044_to_fp16 = const()[name = string("op_2044_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487844160)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_33_cast_fp16 = matmul(transpose_x = c2p_att_33_transpose_x_0, transpose_y = c2p_att_33_transpose_y_0, x = query_layer_51_cast_fp16, y = var_2044_to_fp16)[name = string("c2p_att_33_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_33_cast_fp16 = matmul(transpose_x = c2p_att_33_transpose_x_0, transpose_y = c2p_att_33_transpose_y_0, x = query_layer_51_cast_fp16, y = var_2044_to_fp16)[name = string("c2p_att_33_cast_fp16")];
             bool c2p_att_35_validate_indices_0 = const()[name = string("c2p_att_35_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_35_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_35_validate_indices_0, x = c2p_att_33_cast_fp16)[name = string("c2p_att_35_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_35_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_35_validate_indices_0, x = c2p_att_33_cast_fp16)[name = string("c2p_att_35_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2057_y_0_to_fp16 = const()[name = string("_inversed_2057_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_2057_cast_fp16 = mul(x = c2p_att_35_cast_fp16_cast_int16, y = _inversed_2057_y_0_to_fp16)[name = string("_inversed_2057_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_2057_cast_fp16 = mul(x = c2p_att_35_cast_fp16_cast_int16, y = _inversed_2057_y_0_to_fp16)[name = string("_inversed_2057_cast_fp16")];
             bool p2c_att_33_transpose_x_0 = const()[name = string("p2c_att_33_transpose_x_0"), val = bool(false)];
             bool p2c_att_33_transpose_y_0 = const()[name = string("p2c_att_33_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2072_to_fp16 = const()[name = string("op_2072_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488892800)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_33_cast_fp16 = matmul(transpose_x = p2c_att_33_transpose_x_0, transpose_y = p2c_att_33_transpose_y_0, x = key_layer_17_cast_fp16, y = var_2072_to_fp16)[name = string("p2c_att_33_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_33_cast_fp16 = matmul(transpose_x = p2c_att_33_transpose_x_0, transpose_y = p2c_att_33_transpose_y_0, x = key_layer_17_cast_fp16, y = var_2072_to_fp16)[name = string("p2c_att_33_cast_fp16")];
             bool var_2080_validate_indices_0 = const()[name = string("op_2080_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_2080_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2080_validate_indices_0, x = p2c_att_33_cast_fp16)[name = string("op_2080_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_2080_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2080_validate_indices_0, x = p2c_att_33_cast_fp16)[name = string("op_2080_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_35_perm_0 = const()[name = string("p2c_att_35_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2083_y_0_to_fp16 = const()[name = string("_inversed_2083_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_35_cast_fp16 = transpose(perm = p2c_att_35_perm_0, x = var_2080_cast_fp16_cast_int16)[name = string("transpose_91")];
-            tensor<fp16, [16, 256, 256]> _inversed_2083_cast_fp16 = mul(x = p2c_att_35_cast_fp16, y = _inversed_2083_y_0_to_fp16)[name = string("_inversed_2083_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_17_cast_fp16 = add(x = _inversed_2057_cast_fp16, y = _inversed_2083_cast_fp16)[name = string("rel_att_17_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_51_cast_fp16 = add(x = attention_scores_49_cast_fp16, y = rel_att_17_cast_fp16)[name = string("attention_scores_51_cast_fp16")];
-            tensor<int32, [4]> var_2088 = const()[name = string("op_2088"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_53_cast_fp16 = reshape(shape = var_2088, x = attention_scores_51_cast_fp16)[name = string("attention_scores_53_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_201_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_53_cast_fp16, cond = var_339)[name = string("input_201_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_203_cast_fp16 = softmax(axis = var_56, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
-            tensor<int32, [3]> var_2097 = const()[name = string("op_2097"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_2098_cast_fp16 = reshape(shape = var_2097, x = input_203_cast_fp16)[name = string("op_2098_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_35_cast_fp16 = transpose(perm = p2c_att_35_perm_0, x = var_2080_cast_fp16_cast_int16)[name = string("transpose_91")];
+            tensor<fp16, [16, 192, 192]> _inversed_2083_cast_fp16 = mul(x = p2c_att_35_cast_fp16, y = _inversed_2083_y_0_to_fp16)[name = string("_inversed_2083_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_17_cast_fp16 = add(x = _inversed_2057_cast_fp16, y = _inversed_2083_cast_fp16)[name = string("rel_att_17_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_51_cast_fp16 = add(x = attention_scores_49_cast_fp16, y = rel_att_17_cast_fp16)[name = string("attention_scores_51_cast_fp16")];
+            tensor<int32, [4]> var_2088 = const()[name = string("op_2088"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_53_cast_fp16 = reshape(shape = var_2088, x = attention_scores_51_cast_fp16)[name = string("attention_scores_53_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_201_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_53_cast_fp16, cond = var_339)[name = string("input_201_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_203_cast_fp16 = softmax(axis = var_56, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [3]> var_2097 = const()[name = string("op_2097"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_2098_cast_fp16 = reshape(shape = var_2097, x = input_203_cast_fp16)[name = string("op_2098_cast_fp16")];
             bool context_layer_33_transpose_x_0 = const()[name = string("context_layer_33_transpose_x_0"), val = bool(false)];
             bool context_layer_33_transpose_y_0 = const()[name = string("context_layer_33_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = var_2098_cast_fp16, y = value_layer_17_cast_fp16)[name = string("context_layer_33_cast_fp16")];
-            tensor<int32, [4]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2103_cast_fp16 = reshape(shape = var_2102, x = context_layer_33_cast_fp16)[name = string("op_2103_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = var_2098_cast_fp16, y = value_layer_17_cast_fp16)[name = string("context_layer_33_cast_fp16")];
+            tensor<int32, [4]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2103_cast_fp16 = reshape(shape = var_2102, x = context_layer_33_cast_fp16)[name = string("op_2103_cast_fp16")];
             tensor<int32, [4]> var_2104 = const()[name = string("op_2104"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2109 = const()[name = string("op_2109"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_2105_cast_fp16 = transpose(perm = var_2104, x = var_2103_cast_fp16)[name = string("transpose_90")];
-            tensor<fp16, [1, 256, 1024]> input_205_cast_fp16 = reshape(shape = var_2109, x = var_2105_cast_fp16)[name = string("input_205_cast_fp16")];
+            tensor<int32, [3]> var_2109 = const()[name = string("op_2109"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_2105_cast_fp16 = transpose(perm = var_2104, x = var_2103_cast_fp16)[name = string("transpose_90")];
+            tensor<fp16, [1, 192, 1024]> input_205_cast_fp16 = reshape(shape = var_2109, x = var_2105_cast_fp16)[name = string("input_205_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489941440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492038656)))];
-            tensor<fp16, [1, 256, 1024]> linear_69_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_69_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_197_cast_fp16)[name = string("input_209_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_69_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_69_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_197_cast_fp16)[name = string("input_209_cast_fp16")];
             tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492040768)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492042880)))];
-            tensor<fp16, [1, 256, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_8_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492044992)))];
             tensor<fp16, [4096]> encoder_encoder_layer_8_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500433664)))];
-            tensor<fp16, [1, 256, 4096]> linear_70_cast_fp16 = linear(bias = encoder_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_70_cast_fp16 = linear(bias = encoder_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_70_cast_fp16")];
             string input_215_mode_0 = const()[name = string("input_215_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = string("input_215_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = string("input_215_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_8_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500441920)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508830592)))];
-            tensor<fp16, [1, 256, 1024]> linear_71_cast_fp16 = linear(bias = encoder_encoder_layer_8_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_output_dense_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_71_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_211_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_71_cast_fp16 = linear(bias = encoder_encoder_layer_8_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_output_dense_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_211_cast_fp16)[name = string("input_219_cast_fp16")];
             tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508832704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508834816)))];
-            tensor<fp16, [1, 256, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508836928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510934144)))];
-            tensor<fp16, [1, 256, 1024]> linear_72_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_72_cast_fp16")];
-            tensor<int32, [4]> var_2160 = const()[name = string("op_2160"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_183_cast_fp16 = reshape(shape = var_2160, x = linear_72_cast_fp16)[name = string("x_183_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_72_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_72_cast_fp16")];
+            tensor<int32, [4]> var_2160 = const()[name = string("op_2160"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_183_cast_fp16 = reshape(shape = var_2160, x = linear_72_cast_fp16)[name = string("x_183_cast_fp16")];
             tensor<int32, [4]> var_2162 = const()[name = string("op_2162"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2167 = const()[name = string("op_2167"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2163_cast_fp16 = transpose(perm = var_2162, x = x_183_cast_fp16)[name = string("transpose_89")];
-            tensor<fp16, [16, 256, 64]> query_layer_57_cast_fp16 = reshape(shape = var_2167, x = var_2163_cast_fp16)[name = string("query_layer_57_cast_fp16")];
+            tensor<int32, [3]> var_2167 = const()[name = string("op_2167"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2163_cast_fp16 = transpose(perm = var_2162, x = x_183_cast_fp16)[name = string("transpose_89")];
+            tensor<fp16, [16, 192, 64]> query_layer_57_cast_fp16 = reshape(shape = var_2167, x = var_2163_cast_fp16)[name = string("query_layer_57_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510936256)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513033472)))];
-            tensor<fp16, [1, 256, 1024]> linear_73_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_73_cast_fp16")];
-            tensor<int32, [4]> var_2174 = const()[name = string("op_2174"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_187_cast_fp16 = reshape(shape = var_2174, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_73_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<int32, [4]> var_2174 = const()[name = string("op_2174"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_187_cast_fp16 = reshape(shape = var_2174, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
             tensor<int32, [4]> var_2176 = const()[name = string("op_2176"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2181 = const()[name = string("op_2181"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2177_cast_fp16 = transpose(perm = var_2176, x = x_187_cast_fp16)[name = string("transpose_88")];
-            tensor<fp16, [16, 256, 64]> key_layer_19_cast_fp16 = reshape(shape = var_2181, x = var_2177_cast_fp16)[name = string("key_layer_19_cast_fp16")];
+            tensor<int32, [3]> var_2181 = const()[name = string("op_2181"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2177_cast_fp16 = transpose(perm = var_2176, x = x_187_cast_fp16)[name = string("transpose_88")];
+            tensor<fp16, [16, 192, 64]> key_layer_19_cast_fp16 = reshape(shape = var_2181, x = var_2177_cast_fp16)[name = string("key_layer_19_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513035584)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515132800)))];
-            tensor<fp16, [1, 256, 1024]> linear_74_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_74_cast_fp16")];
-            tensor<int32, [4]> var_2188 = const()[name = string("op_2188"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_191_cast_fp16 = reshape(shape = var_2188, x = linear_74_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_74_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_74_cast_fp16")];
+            tensor<int32, [4]> var_2188 = const()[name = string("op_2188"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_191_cast_fp16 = reshape(shape = var_2188, x = linear_74_cast_fp16)[name = string("x_191_cast_fp16")];
             tensor<int32, [4]> var_2190 = const()[name = string("op_2190"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2195 = const()[name = string("op_2195"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2191_cast_fp16 = transpose(perm = var_2190, x = x_191_cast_fp16)[name = string("transpose_87")];
-            tensor<fp16, [16, 256, 64]> value_layer_19_cast_fp16 = reshape(shape = var_2195, x = var_2191_cast_fp16)[name = string("value_layer_19_cast_fp16")];
+            tensor<int32, [3]> var_2195 = const()[name = string("op_2195"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2191_cast_fp16 = transpose(perm = var_2190, x = x_191_cast_fp16)[name = string("transpose_87")];
+            tensor<fp16, [16, 192, 64]> value_layer_19_cast_fp16 = reshape(shape = var_2195, x = var_2191_cast_fp16)[name = string("value_layer_19_cast_fp16")];
             tensor<int32, [3]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2207_y_0_to_fp16 = const()[name = string("_inversed_2207_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = key_layer_19_cast_fp16)[name = string("transpose_86")];
-            tensor<fp16, [16, 64, 256]> _inversed_2207_cast_fp16 = mul(x = var_2205_cast_fp16, y = _inversed_2207_y_0_to_fp16)[name = string("_inversed_2207_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = key_layer_19_cast_fp16)[name = string("transpose_86")];
+            tensor<fp16, [16, 64, 192]> _inversed_2207_cast_fp16 = mul(x = var_2205_cast_fp16, y = _inversed_2207_y_0_to_fp16)[name = string("_inversed_2207_cast_fp16")];
             bool attention_scores_55_transpose_x_0 = const()[name = string("attention_scores_55_transpose_x_0"), val = bool(false)];
             bool attention_scores_55_transpose_y_0 = const()[name = string("attention_scores_55_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_55_cast_fp16 = matmul(transpose_x = attention_scores_55_transpose_x_0, transpose_y = attention_scores_55_transpose_y_0, x = query_layer_57_cast_fp16, y = _inversed_2207_cast_fp16)[name = string("attention_scores_55_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_55_cast_fp16 = matmul(transpose_x = attention_scores_55_transpose_x_0, transpose_y = attention_scores_55_transpose_y_0, x = query_layer_57_cast_fp16, y = _inversed_2207_cast_fp16)[name = string("attention_scores_55_cast_fp16")];
             bool c2p_att_37_transpose_x_0 = const()[name = string("c2p_att_37_transpose_x_0"), val = bool(false)];
             bool c2p_att_37_transpose_y_0 = const()[name = string("c2p_att_37_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2263_to_fp16 = const()[name = string("op_2263_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515134912)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_37_cast_fp16 = matmul(transpose_x = c2p_att_37_transpose_x_0, transpose_y = c2p_att_37_transpose_y_0, x = query_layer_57_cast_fp16, y = var_2263_to_fp16)[name = string("c2p_att_37_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_37_cast_fp16 = matmul(transpose_x = c2p_att_37_transpose_x_0, transpose_y = c2p_att_37_transpose_y_0, x = query_layer_57_cast_fp16, y = var_2263_to_fp16)[name = string("c2p_att_37_cast_fp16")];
             bool c2p_att_39_validate_indices_0 = const()[name = string("c2p_att_39_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_39_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_39_validate_indices_0, x = c2p_att_37_cast_fp16)[name = string("c2p_att_39_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_39_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_39_validate_indices_0, x = c2p_att_37_cast_fp16)[name = string("c2p_att_39_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2276_y_0_to_fp16 = const()[name = string("_inversed_2276_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_2276_cast_fp16 = mul(x = c2p_att_39_cast_fp16_cast_int16, y = _inversed_2276_y_0_to_fp16)[name = string("_inversed_2276_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_2276_cast_fp16 = mul(x = c2p_att_39_cast_fp16_cast_int16, y = _inversed_2276_y_0_to_fp16)[name = string("_inversed_2276_cast_fp16")];
             bool p2c_att_37_transpose_x_0 = const()[name = string("p2c_att_37_transpose_x_0"), val = bool(false)];
             bool p2c_att_37_transpose_y_0 = const()[name = string("p2c_att_37_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2291_to_fp16 = const()[name = string("op_2291_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516183552)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_37_cast_fp16 = matmul(transpose_x = p2c_att_37_transpose_x_0, transpose_y = p2c_att_37_transpose_y_0, x = key_layer_19_cast_fp16, y = var_2291_to_fp16)[name = string("p2c_att_37_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_37_cast_fp16 = matmul(transpose_x = p2c_att_37_transpose_x_0, transpose_y = p2c_att_37_transpose_y_0, x = key_layer_19_cast_fp16, y = var_2291_to_fp16)[name = string("p2c_att_37_cast_fp16")];
             bool var_2299_validate_indices_0 = const()[name = string("op_2299_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_2299_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2299_validate_indices_0, x = p2c_att_37_cast_fp16)[name = string("op_2299_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_2299_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2299_validate_indices_0, x = p2c_att_37_cast_fp16)[name = string("op_2299_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_39_perm_0 = const()[name = string("p2c_att_39_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2302_y_0_to_fp16 = const()[name = string("_inversed_2302_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_39_cast_fp16 = transpose(perm = p2c_att_39_perm_0, x = var_2299_cast_fp16_cast_int16)[name = string("transpose_85")];
-            tensor<fp16, [16, 256, 256]> _inversed_2302_cast_fp16 = mul(x = p2c_att_39_cast_fp16, y = _inversed_2302_y_0_to_fp16)[name = string("_inversed_2302_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_19_cast_fp16 = add(x = _inversed_2276_cast_fp16, y = _inversed_2302_cast_fp16)[name = string("rel_att_19_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_57_cast_fp16 = add(x = attention_scores_55_cast_fp16, y = rel_att_19_cast_fp16)[name = string("attention_scores_57_cast_fp16")];
-            tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_59_cast_fp16 = reshape(shape = var_2307, x = attention_scores_57_cast_fp16)[name = string("attention_scores_59_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_225_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_59_cast_fp16, cond = var_339)[name = string("input_225_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_227_cast_fp16 = softmax(axis = var_56, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
-            tensor<int32, [3]> var_2316 = const()[name = string("op_2316"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_2317_cast_fp16 = reshape(shape = var_2316, x = input_227_cast_fp16)[name = string("op_2317_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_39_cast_fp16 = transpose(perm = p2c_att_39_perm_0, x = var_2299_cast_fp16_cast_int16)[name = string("transpose_85")];
+            tensor<fp16, [16, 192, 192]> _inversed_2302_cast_fp16 = mul(x = p2c_att_39_cast_fp16, y = _inversed_2302_y_0_to_fp16)[name = string("_inversed_2302_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_19_cast_fp16 = add(x = _inversed_2276_cast_fp16, y = _inversed_2302_cast_fp16)[name = string("rel_att_19_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_57_cast_fp16 = add(x = attention_scores_55_cast_fp16, y = rel_att_19_cast_fp16)[name = string("attention_scores_57_cast_fp16")];
+            tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_59_cast_fp16 = reshape(shape = var_2307, x = attention_scores_57_cast_fp16)[name = string("attention_scores_59_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_225_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_59_cast_fp16, cond = var_339)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_227_cast_fp16 = softmax(axis = var_56, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
+            tensor<int32, [3]> var_2316 = const()[name = string("op_2316"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_2317_cast_fp16 = reshape(shape = var_2316, x = input_227_cast_fp16)[name = string("op_2317_cast_fp16")];
             bool context_layer_37_transpose_x_0 = const()[name = string("context_layer_37_transpose_x_0"), val = bool(false)];
             bool context_layer_37_transpose_y_0 = const()[name = string("context_layer_37_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = var_2317_cast_fp16, y = value_layer_19_cast_fp16)[name = string("context_layer_37_cast_fp16")];
-            tensor<int32, [4]> var_2321 = const()[name = string("op_2321"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2322_cast_fp16 = reshape(shape = var_2321, x = context_layer_37_cast_fp16)[name = string("op_2322_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = var_2317_cast_fp16, y = value_layer_19_cast_fp16)[name = string("context_layer_37_cast_fp16")];
+            tensor<int32, [4]> var_2321 = const()[name = string("op_2321"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2322_cast_fp16 = reshape(shape = var_2321, x = context_layer_37_cast_fp16)[name = string("op_2322_cast_fp16")];
             tensor<int32, [4]> var_2323 = const()[name = string("op_2323"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_2324_cast_fp16 = transpose(perm = var_2323, x = var_2322_cast_fp16)[name = string("transpose_84")];
-            tensor<fp16, [1, 256, 1024]> input_229_cast_fp16 = reshape(shape = var_2328, x = var_2324_cast_fp16)[name = string("input_229_cast_fp16")];
+            tensor<int32, [3]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_2324_cast_fp16 = transpose(perm = var_2323, x = var_2322_cast_fp16)[name = string("transpose_84")];
+            tensor<fp16, [1, 192, 1024]> input_229_cast_fp16 = reshape(shape = var_2328, x = var_2324_cast_fp16)[name = string("input_229_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517232192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519329408)))];
-            tensor<fp16, [1, 256, 1024]> linear_77_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_77_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_233_cast_fp16 = add(x = linear_77_cast_fp16, y = input_221_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_77_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_233_cast_fp16 = add(x = linear_77_cast_fp16, y = input_221_cast_fp16)[name = string("input_233_cast_fp16")];
             tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519331520)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519333632)))];
-            tensor<fp16, [1, 256, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_9_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519335744)))];
             tensor<fp16, [4096]> encoder_encoder_layer_9_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527724416)))];
-            tensor<fp16, [1, 256, 4096]> linear_78_cast_fp16 = linear(bias = encoder_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_78_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_78_cast_fp16 = linear(bias = encoder_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_78_cast_fp16")];
             string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = string("input_239_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_9_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527732672)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536121344)))];
-            tensor<fp16, [1, 256, 1024]> linear_79_cast_fp16 = linear(bias = encoder_encoder_layer_9_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_output_dense_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_79_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_243_cast_fp16 = add(x = linear_79_cast_fp16, y = input_235_cast_fp16)[name = string("input_243_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_79_cast_fp16 = linear(bias = encoder_encoder_layer_9_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_output_dense_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_243_cast_fp16 = add(x = linear_79_cast_fp16, y = input_235_cast_fp16)[name = string("input_243_cast_fp16")];
             tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536123456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536125568)))];
-            tensor<fp16, [1, 256, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536127680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538224896)))];
-            tensor<fp16, [1, 256, 1024]> linear_80_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_80_cast_fp16")];
-            tensor<int32, [4]> var_2379 = const()[name = string("op_2379"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_203_cast_fp16 = reshape(shape = var_2379, x = linear_80_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_80_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<int32, [4]> var_2379 = const()[name = string("op_2379"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_203_cast_fp16 = reshape(shape = var_2379, x = linear_80_cast_fp16)[name = string("x_203_cast_fp16")];
             tensor<int32, [4]> var_2381 = const()[name = string("op_2381"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2382_cast_fp16 = transpose(perm = var_2381, x = x_203_cast_fp16)[name = string("transpose_83")];
-            tensor<fp16, [16, 256, 64]> query_layer_63_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("query_layer_63_cast_fp16")];
+            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2382_cast_fp16 = transpose(perm = var_2381, x = x_203_cast_fp16)[name = string("transpose_83")];
+            tensor<fp16, [16, 192, 64]> query_layer_63_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("query_layer_63_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538227008)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540324224)))];
-            tensor<fp16, [1, 256, 1024]> linear_81_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_81_cast_fp16")];
-            tensor<int32, [4]> var_2393 = const()[name = string("op_2393"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_207_cast_fp16 = reshape(shape = var_2393, x = linear_81_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_81_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_81_cast_fp16")];
+            tensor<int32, [4]> var_2393 = const()[name = string("op_2393"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_207_cast_fp16 = reshape(shape = var_2393, x = linear_81_cast_fp16)[name = string("x_207_cast_fp16")];
             tensor<int32, [4]> var_2395 = const()[name = string("op_2395"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2400 = const()[name = string("op_2400"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2396_cast_fp16 = transpose(perm = var_2395, x = x_207_cast_fp16)[name = string("transpose_82")];
-            tensor<fp16, [16, 256, 64]> key_layer_21_cast_fp16 = reshape(shape = var_2400, x = var_2396_cast_fp16)[name = string("key_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_2400 = const()[name = string("op_2400"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2396_cast_fp16 = transpose(perm = var_2395, x = x_207_cast_fp16)[name = string("transpose_82")];
+            tensor<fp16, [16, 192, 64]> key_layer_21_cast_fp16 = reshape(shape = var_2400, x = var_2396_cast_fp16)[name = string("key_layer_21_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540326336)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542423552)))];
-            tensor<fp16, [1, 256, 1024]> linear_82_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_82_cast_fp16")];
-            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_211_cast_fp16 = reshape(shape = var_2407, x = linear_82_cast_fp16)[name = string("x_211_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_82_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_211_cast_fp16 = reshape(shape = var_2407, x = linear_82_cast_fp16)[name = string("x_211_cast_fp16")];
             tensor<int32, [4]> var_2409 = const()[name = string("op_2409"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2414 = const()[name = string("op_2414"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2410_cast_fp16 = transpose(perm = var_2409, x = x_211_cast_fp16)[name = string("transpose_81")];
-            tensor<fp16, [16, 256, 64]> value_layer_21_cast_fp16 = reshape(shape = var_2414, x = var_2410_cast_fp16)[name = string("value_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_2414 = const()[name = string("op_2414"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2410_cast_fp16 = transpose(perm = var_2409, x = x_211_cast_fp16)[name = string("transpose_81")];
+            tensor<fp16, [16, 192, 64]> value_layer_21_cast_fp16 = reshape(shape = var_2414, x = var_2410_cast_fp16)[name = string("value_layer_21_cast_fp16")];
             tensor<int32, [3]> var_2424_perm_0 = const()[name = string("op_2424_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2426_y_0_to_fp16 = const()[name = string("_inversed_2426_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_2424_cast_fp16 = transpose(perm = var_2424_perm_0, x = key_layer_21_cast_fp16)[name = string("transpose_80")];
-            tensor<fp16, [16, 64, 256]> _inversed_2426_cast_fp16 = mul(x = var_2424_cast_fp16, y = _inversed_2426_y_0_to_fp16)[name = string("_inversed_2426_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_2424_cast_fp16 = transpose(perm = var_2424_perm_0, x = key_layer_21_cast_fp16)[name = string("transpose_80")];
+            tensor<fp16, [16, 64, 192]> _inversed_2426_cast_fp16 = mul(x = var_2424_cast_fp16, y = _inversed_2426_y_0_to_fp16)[name = string("_inversed_2426_cast_fp16")];
             bool attention_scores_61_transpose_x_0 = const()[name = string("attention_scores_61_transpose_x_0"), val = bool(false)];
             bool attention_scores_61_transpose_y_0 = const()[name = string("attention_scores_61_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_61_cast_fp16 = matmul(transpose_x = attention_scores_61_transpose_x_0, transpose_y = attention_scores_61_transpose_y_0, x = query_layer_63_cast_fp16, y = _inversed_2426_cast_fp16)[name = string("attention_scores_61_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_61_cast_fp16 = matmul(transpose_x = attention_scores_61_transpose_x_0, transpose_y = attention_scores_61_transpose_y_0, x = query_layer_63_cast_fp16, y = _inversed_2426_cast_fp16)[name = string("attention_scores_61_cast_fp16")];
             bool c2p_att_41_transpose_x_0 = const()[name = string("c2p_att_41_transpose_x_0"), val = bool(false)];
             bool c2p_att_41_transpose_y_0 = const()[name = string("c2p_att_41_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542425664)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_41_cast_fp16 = matmul(transpose_x = c2p_att_41_transpose_x_0, transpose_y = c2p_att_41_transpose_y_0, x = query_layer_63_cast_fp16, y = var_2482_to_fp16)[name = string("c2p_att_41_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_41_cast_fp16 = matmul(transpose_x = c2p_att_41_transpose_x_0, transpose_y = c2p_att_41_transpose_y_0, x = query_layer_63_cast_fp16, y = var_2482_to_fp16)[name = string("c2p_att_41_cast_fp16")];
             bool c2p_att_43_validate_indices_0 = const()[name = string("c2p_att_43_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_43_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_43_validate_indices_0, x = c2p_att_41_cast_fp16)[name = string("c2p_att_43_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_43_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_43_validate_indices_0, x = c2p_att_41_cast_fp16)[name = string("c2p_att_43_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2495_y_0_to_fp16 = const()[name = string("_inversed_2495_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_2495_cast_fp16 = mul(x = c2p_att_43_cast_fp16_cast_int16, y = _inversed_2495_y_0_to_fp16)[name = string("_inversed_2495_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_2495_cast_fp16 = mul(x = c2p_att_43_cast_fp16_cast_int16, y = _inversed_2495_y_0_to_fp16)[name = string("_inversed_2495_cast_fp16")];
             bool p2c_att_41_transpose_x_0 = const()[name = string("p2c_att_41_transpose_x_0"), val = bool(false)];
             bool p2c_att_41_transpose_y_0 = const()[name = string("p2c_att_41_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543474304)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_41_cast_fp16 = matmul(transpose_x = p2c_att_41_transpose_x_0, transpose_y = p2c_att_41_transpose_y_0, x = key_layer_21_cast_fp16, y = var_2510_to_fp16)[name = string("p2c_att_41_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_41_cast_fp16 = matmul(transpose_x = p2c_att_41_transpose_x_0, transpose_y = p2c_att_41_transpose_y_0, x = key_layer_21_cast_fp16, y = var_2510_to_fp16)[name = string("p2c_att_41_cast_fp16")];
             bool var_2518_validate_indices_0 = const()[name = string("op_2518_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_2518_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2518_validate_indices_0, x = p2c_att_41_cast_fp16)[name = string("op_2518_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_2518_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2518_validate_indices_0, x = p2c_att_41_cast_fp16)[name = string("op_2518_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_43_perm_0 = const()[name = string("p2c_att_43_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2521_y_0_to_fp16 = const()[name = string("_inversed_2521_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_43_cast_fp16 = transpose(perm = p2c_att_43_perm_0, x = var_2518_cast_fp16_cast_int16)[name = string("transpose_79")];
-            tensor<fp16, [16, 256, 256]> _inversed_2521_cast_fp16 = mul(x = p2c_att_43_cast_fp16, y = _inversed_2521_y_0_to_fp16)[name = string("_inversed_2521_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_21_cast_fp16 = add(x = _inversed_2495_cast_fp16, y = _inversed_2521_cast_fp16)[name = string("rel_att_21_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_63_cast_fp16 = add(x = attention_scores_61_cast_fp16, y = rel_att_21_cast_fp16)[name = string("attention_scores_63_cast_fp16")];
-            tensor<int32, [4]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_65_cast_fp16 = reshape(shape = var_2526, x = attention_scores_63_cast_fp16)[name = string("attention_scores_65_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_249_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_65_cast_fp16, cond = var_339)[name = string("input_249_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_251_cast_fp16 = softmax(axis = var_56, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
-            tensor<int32, [3]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_2536_cast_fp16 = reshape(shape = var_2535, x = input_251_cast_fp16)[name = string("op_2536_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_43_cast_fp16 = transpose(perm = p2c_att_43_perm_0, x = var_2518_cast_fp16_cast_int16)[name = string("transpose_79")];
+            tensor<fp16, [16, 192, 192]> _inversed_2521_cast_fp16 = mul(x = p2c_att_43_cast_fp16, y = _inversed_2521_y_0_to_fp16)[name = string("_inversed_2521_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_21_cast_fp16 = add(x = _inversed_2495_cast_fp16, y = _inversed_2521_cast_fp16)[name = string("rel_att_21_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_63_cast_fp16 = add(x = attention_scores_61_cast_fp16, y = rel_att_21_cast_fp16)[name = string("attention_scores_63_cast_fp16")];
+            tensor<int32, [4]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_65_cast_fp16 = reshape(shape = var_2526, x = attention_scores_63_cast_fp16)[name = string("attention_scores_65_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_249_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_65_cast_fp16, cond = var_339)[name = string("input_249_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_251_cast_fp16 = softmax(axis = var_56, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<int32, [3]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_2536_cast_fp16 = reshape(shape = var_2535, x = input_251_cast_fp16)[name = string("op_2536_cast_fp16")];
             bool context_layer_41_transpose_x_0 = const()[name = string("context_layer_41_transpose_x_0"), val = bool(false)];
             bool context_layer_41_transpose_y_0 = const()[name = string("context_layer_41_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = var_2536_cast_fp16, y = value_layer_21_cast_fp16)[name = string("context_layer_41_cast_fp16")];
-            tensor<int32, [4]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2541_cast_fp16 = reshape(shape = var_2540, x = context_layer_41_cast_fp16)[name = string("op_2541_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = var_2536_cast_fp16, y = value_layer_21_cast_fp16)[name = string("context_layer_41_cast_fp16")];
+            tensor<int32, [4]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2541_cast_fp16 = reshape(shape = var_2540, x = context_layer_41_cast_fp16)[name = string("op_2541_cast_fp16")];
             tensor<int32, [4]> var_2542 = const()[name = string("op_2542"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_2543_cast_fp16 = transpose(perm = var_2542, x = var_2541_cast_fp16)[name = string("transpose_78")];
-            tensor<fp16, [1, 256, 1024]> input_253_cast_fp16 = reshape(shape = var_2547, x = var_2543_cast_fp16)[name = string("input_253_cast_fp16")];
+            tensor<int32, [3]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_2543_cast_fp16 = transpose(perm = var_2542, x = var_2541_cast_fp16)[name = string("transpose_78")];
+            tensor<fp16, [1, 192, 1024]> input_253_cast_fp16 = reshape(shape = var_2547, x = var_2543_cast_fp16)[name = string("input_253_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544522944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546620160)))];
-            tensor<fp16, [1, 256, 1024]> linear_85_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_253_cast_fp16)[name = string("linear_85_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_257_cast_fp16 = add(x = linear_85_cast_fp16, y = input_245_cast_fp16)[name = string("input_257_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_85_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_253_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_257_cast_fp16 = add(x = linear_85_cast_fp16, y = input_245_cast_fp16)[name = string("input_257_cast_fp16")];
             tensor<int32, [1]> input_259_axes_0 = const()[name = string("input_259_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546622272)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546624384)))];
-            tensor<fp16, [1, 256, 1024]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_10_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546626496)))];
             tensor<fp16, [4096]> encoder_encoder_layer_10_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555015168)))];
-            tensor<fp16, [1, 256, 4096]> linear_86_cast_fp16 = linear(bias = encoder_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_86_cast_fp16 = linear(bias = encoder_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_86_cast_fp16")];
             string input_263_mode_0 = const()[name = string("input_263_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = string("input_263_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = string("input_263_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_10_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555023424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563412096)))];
-            tensor<fp16, [1, 256, 1024]> linear_87_cast_fp16 = linear(bias = encoder_encoder_layer_10_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_output_dense_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_87_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_267_cast_fp16 = add(x = linear_87_cast_fp16, y = input_259_cast_fp16)[name = string("input_267_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_87_cast_fp16 = linear(bias = encoder_encoder_layer_10_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_output_dense_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_87_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_267_cast_fp16 = add(x = linear_87_cast_fp16, y = input_259_cast_fp16)[name = string("input_267_cast_fp16")];
             tensor<int32, [1]> input_269_axes_0 = const()[name = string("input_269_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563414208)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563416320)))];
-            tensor<fp16, [1, 256, 1024]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563418432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565515648)))];
-            tensor<fp16, [1, 256, 1024]> linear_88_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_88_cast_fp16")];
-            tensor<int32, [4]> var_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_223_cast_fp16 = reshape(shape = var_2598, x = linear_88_cast_fp16)[name = string("x_223_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_88_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<int32, [4]> var_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_223_cast_fp16 = reshape(shape = var_2598, x = linear_88_cast_fp16)[name = string("x_223_cast_fp16")];
             tensor<int32, [4]> var_2600 = const()[name = string("op_2600"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2601_cast_fp16 = transpose(perm = var_2600, x = x_223_cast_fp16)[name = string("transpose_77")];
-            tensor<fp16, [16, 256, 64]> query_layer_69_cast_fp16 = reshape(shape = var_2605, x = var_2601_cast_fp16)[name = string("query_layer_69_cast_fp16")];
+            tensor<int32, [3]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2601_cast_fp16 = transpose(perm = var_2600, x = x_223_cast_fp16)[name = string("transpose_77")];
+            tensor<fp16, [16, 192, 64]> query_layer_69_cast_fp16 = reshape(shape = var_2605, x = var_2601_cast_fp16)[name = string("query_layer_69_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565517760)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567614976)))];
-            tensor<fp16, [1, 256, 1024]> linear_89_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_89_cast_fp16")];
-            tensor<int32, [4]> var_2612 = const()[name = string("op_2612"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_227_cast_fp16 = reshape(shape = var_2612, x = linear_89_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_89_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<int32, [4]> var_2612 = const()[name = string("op_2612"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_227_cast_fp16 = reshape(shape = var_2612, x = linear_89_cast_fp16)[name = string("x_227_cast_fp16")];
             tensor<int32, [4]> var_2614 = const()[name = string("op_2614"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2615_cast_fp16 = transpose(perm = var_2614, x = x_227_cast_fp16)[name = string("transpose_76")];
-            tensor<fp16, [16, 256, 64]> key_layer_23_cast_fp16 = reshape(shape = var_2619, x = var_2615_cast_fp16)[name = string("key_layer_23_cast_fp16")];
+            tensor<int32, [3]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2615_cast_fp16 = transpose(perm = var_2614, x = x_227_cast_fp16)[name = string("transpose_76")];
+            tensor<fp16, [16, 192, 64]> key_layer_23_cast_fp16 = reshape(shape = var_2619, x = var_2615_cast_fp16)[name = string("key_layer_23_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567617088)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569714304)))];
-            tensor<fp16, [1, 256, 1024]> linear_90_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_90_cast_fp16")];
-            tensor<int32, [4]> var_2626 = const()[name = string("op_2626"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_231_cast_fp16 = reshape(shape = var_2626, x = linear_90_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_90_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_90_cast_fp16")];
+            tensor<int32, [4]> var_2626 = const()[name = string("op_2626"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_231_cast_fp16 = reshape(shape = var_2626, x = linear_90_cast_fp16)[name = string("x_231_cast_fp16")];
             tensor<int32, [4]> var_2628 = const()[name = string("op_2628"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2633 = const()[name = string("op_2633"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2629_cast_fp16 = transpose(perm = var_2628, x = x_231_cast_fp16)[name = string("transpose_75")];
-            tensor<fp16, [16, 256, 64]> value_layer_23_cast_fp16 = reshape(shape = var_2633, x = var_2629_cast_fp16)[name = string("value_layer_23_cast_fp16")];
+            tensor<int32, [3]> var_2633 = const()[name = string("op_2633"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2629_cast_fp16 = transpose(perm = var_2628, x = x_231_cast_fp16)[name = string("transpose_75")];
+            tensor<fp16, [16, 192, 64]> value_layer_23_cast_fp16 = reshape(shape = var_2633, x = var_2629_cast_fp16)[name = string("value_layer_23_cast_fp16")];
             tensor<int32, [3]> var_2643_perm_0 = const()[name = string("op_2643_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2645_y_0_to_fp16 = const()[name = string("_inversed_2645_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_2643_cast_fp16 = transpose(perm = var_2643_perm_0, x = key_layer_23_cast_fp16)[name = string("transpose_74")];
-            tensor<fp16, [16, 64, 256]> _inversed_2645_cast_fp16 = mul(x = var_2643_cast_fp16, y = _inversed_2645_y_0_to_fp16)[name = string("_inversed_2645_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_2643_cast_fp16 = transpose(perm = var_2643_perm_0, x = key_layer_23_cast_fp16)[name = string("transpose_74")];
+            tensor<fp16, [16, 64, 192]> _inversed_2645_cast_fp16 = mul(x = var_2643_cast_fp16, y = _inversed_2645_y_0_to_fp16)[name = string("_inversed_2645_cast_fp16")];
             bool attention_scores_67_transpose_x_0 = const()[name = string("attention_scores_67_transpose_x_0"), val = bool(false)];
             bool attention_scores_67_transpose_y_0 = const()[name = string("attention_scores_67_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_67_cast_fp16 = matmul(transpose_x = attention_scores_67_transpose_x_0, transpose_y = attention_scores_67_transpose_y_0, x = query_layer_69_cast_fp16, y = _inversed_2645_cast_fp16)[name = string("attention_scores_67_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_67_cast_fp16 = matmul(transpose_x = attention_scores_67_transpose_x_0, transpose_y = attention_scores_67_transpose_y_0, x = query_layer_69_cast_fp16, y = _inversed_2645_cast_fp16)[name = string("attention_scores_67_cast_fp16")];
             bool c2p_att_45_transpose_x_0 = const()[name = string("c2p_att_45_transpose_x_0"), val = bool(false)];
             bool c2p_att_45_transpose_y_0 = const()[name = string("c2p_att_45_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2701_to_fp16 = const()[name = string("op_2701_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569716416)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_45_cast_fp16 = matmul(transpose_x = c2p_att_45_transpose_x_0, transpose_y = c2p_att_45_transpose_y_0, x = query_layer_69_cast_fp16, y = var_2701_to_fp16)[name = string("c2p_att_45_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_45_cast_fp16 = matmul(transpose_x = c2p_att_45_transpose_x_0, transpose_y = c2p_att_45_transpose_y_0, x = query_layer_69_cast_fp16, y = var_2701_to_fp16)[name = string("c2p_att_45_cast_fp16")];
             bool c2p_att_47_validate_indices_0 = const()[name = string("c2p_att_47_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_47_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_47_validate_indices_0, x = c2p_att_45_cast_fp16)[name = string("c2p_att_47_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_47_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_47_validate_indices_0, x = c2p_att_45_cast_fp16)[name = string("c2p_att_47_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2714_y_0_to_fp16 = const()[name = string("_inversed_2714_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_2714_cast_fp16 = mul(x = c2p_att_47_cast_fp16_cast_int16, y = _inversed_2714_y_0_to_fp16)[name = string("_inversed_2714_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_2714_cast_fp16 = mul(x = c2p_att_47_cast_fp16_cast_int16, y = _inversed_2714_y_0_to_fp16)[name = string("_inversed_2714_cast_fp16")];
             bool p2c_att_45_transpose_x_0 = const()[name = string("p2c_att_45_transpose_x_0"), val = bool(false)];
             bool p2c_att_45_transpose_y_0 = const()[name = string("p2c_att_45_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2729_to_fp16 = const()[name = string("op_2729_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570765056)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_45_cast_fp16 = matmul(transpose_x = p2c_att_45_transpose_x_0, transpose_y = p2c_att_45_transpose_y_0, x = key_layer_23_cast_fp16, y = var_2729_to_fp16)[name = string("p2c_att_45_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_45_cast_fp16 = matmul(transpose_x = p2c_att_45_transpose_x_0, transpose_y = p2c_att_45_transpose_y_0, x = key_layer_23_cast_fp16, y = var_2729_to_fp16)[name = string("p2c_att_45_cast_fp16")];
             bool var_2737_validate_indices_0 = const()[name = string("op_2737_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_2737_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2737_validate_indices_0, x = p2c_att_45_cast_fp16)[name = string("op_2737_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_2737_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2737_validate_indices_0, x = p2c_att_45_cast_fp16)[name = string("op_2737_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_47_perm_0 = const()[name = string("p2c_att_47_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2740_y_0_to_fp16 = const()[name = string("_inversed_2740_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_47_cast_fp16 = transpose(perm = p2c_att_47_perm_0, x = var_2737_cast_fp16_cast_int16)[name = string("transpose_73")];
-            tensor<fp16, [16, 256, 256]> _inversed_2740_cast_fp16 = mul(x = p2c_att_47_cast_fp16, y = _inversed_2740_y_0_to_fp16)[name = string("_inversed_2740_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_23_cast_fp16 = add(x = _inversed_2714_cast_fp16, y = _inversed_2740_cast_fp16)[name = string("rel_att_23_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_69_cast_fp16 = add(x = attention_scores_67_cast_fp16, y = rel_att_23_cast_fp16)[name = string("attention_scores_69_cast_fp16")];
-            tensor<int32, [4]> var_2745 = const()[name = string("op_2745"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_71_cast_fp16 = reshape(shape = var_2745, x = attention_scores_69_cast_fp16)[name = string("attention_scores_71_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_273_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_71_cast_fp16, cond = var_339)[name = string("input_273_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_275_cast_fp16 = softmax(axis = var_56, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")];
-            tensor<int32, [3]> var_2754 = const()[name = string("op_2754"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_2755_cast_fp16 = reshape(shape = var_2754, x = input_275_cast_fp16)[name = string("op_2755_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_47_cast_fp16 = transpose(perm = p2c_att_47_perm_0, x = var_2737_cast_fp16_cast_int16)[name = string("transpose_73")];
+            tensor<fp16, [16, 192, 192]> _inversed_2740_cast_fp16 = mul(x = p2c_att_47_cast_fp16, y = _inversed_2740_y_0_to_fp16)[name = string("_inversed_2740_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_23_cast_fp16 = add(x = _inversed_2714_cast_fp16, y = _inversed_2740_cast_fp16)[name = string("rel_att_23_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_69_cast_fp16 = add(x = attention_scores_67_cast_fp16, y = rel_att_23_cast_fp16)[name = string("attention_scores_69_cast_fp16")];
+            tensor<int32, [4]> var_2745 = const()[name = string("op_2745"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_71_cast_fp16 = reshape(shape = var_2745, x = attention_scores_69_cast_fp16)[name = string("attention_scores_71_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_273_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_71_cast_fp16, cond = var_339)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_275_cast_fp16 = softmax(axis = var_56, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")];
+            tensor<int32, [3]> var_2754 = const()[name = string("op_2754"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_2755_cast_fp16 = reshape(shape = var_2754, x = input_275_cast_fp16)[name = string("op_2755_cast_fp16")];
             bool context_layer_45_transpose_x_0 = const()[name = string("context_layer_45_transpose_x_0"), val = bool(false)];
             bool context_layer_45_transpose_y_0 = const()[name = string("context_layer_45_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = var_2755_cast_fp16, y = value_layer_23_cast_fp16)[name = string("context_layer_45_cast_fp16")];
-            tensor<int32, [4]> var_2759 = const()[name = string("op_2759"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2760_cast_fp16 = reshape(shape = var_2759, x = context_layer_45_cast_fp16)[name = string("op_2760_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = var_2755_cast_fp16, y = value_layer_23_cast_fp16)[name = string("context_layer_45_cast_fp16")];
+            tensor<int32, [4]> var_2759 = const()[name = string("op_2759"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2760_cast_fp16 = reshape(shape = var_2759, x = context_layer_45_cast_fp16)[name = string("op_2760_cast_fp16")];
             tensor<int32, [4]> var_2761 = const()[name = string("op_2761"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2766 = const()[name = string("op_2766"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_2762_cast_fp16 = transpose(perm = var_2761, x = var_2760_cast_fp16)[name = string("transpose_72")];
-            tensor<fp16, [1, 256, 1024]> input_277_cast_fp16 = reshape(shape = var_2766, x = var_2762_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<int32, [3]> var_2766 = const()[name = string("op_2766"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_2762_cast_fp16 = transpose(perm = var_2761, x = var_2760_cast_fp16)[name = string("transpose_72")];
+            tensor<fp16, [1, 192, 1024]> input_277_cast_fp16 = reshape(shape = var_2766, x = var_2762_cast_fp16)[name = string("input_277_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571813696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573910912)))];
-            tensor<fp16, [1, 256, 1024]> linear_93_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_277_cast_fp16)[name = string("linear_93_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_281_cast_fp16 = add(x = linear_93_cast_fp16, y = input_269_cast_fp16)[name = string("input_281_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_93_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_277_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_281_cast_fp16 = add(x = linear_93_cast_fp16, y = input_269_cast_fp16)[name = string("input_281_cast_fp16")];
             tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573913024)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573915136)))];
-            tensor<fp16, [1, 256, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_11_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573917248)))];
             tensor<fp16, [4096]> encoder_encoder_layer_11_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582305920)))];
-            tensor<fp16, [1, 256, 4096]> linear_94_cast_fp16 = linear(bias = encoder_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_94_cast_fp16 = linear(bias = encoder_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_94_cast_fp16")];
             string input_287_mode_0 = const()[name = string("input_287_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_287_cast_fp16 = gelu(mode = input_287_mode_0, x = linear_94_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_287_cast_fp16 = gelu(mode = input_287_mode_0, x = linear_94_cast_fp16)[name = string("input_287_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_11_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582314176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590702848)))];
-            tensor<fp16, [1, 256, 1024]> linear_95_cast_fp16 = linear(bias = encoder_encoder_layer_11_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_output_dense_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_95_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_291_cast_fp16 = add(x = linear_95_cast_fp16, y = input_283_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_95_cast_fp16 = linear(bias = encoder_encoder_layer_11_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_output_dense_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_291_cast_fp16 = add(x = linear_95_cast_fp16, y = input_283_cast_fp16)[name = string("input_291_cast_fp16")];
             tensor<int32, [1]> input_293_axes_0 = const()[name = string("input_293_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590704960)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590707072)))];
-            tensor<fp16, [1, 256, 1024]> input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_291_cast_fp16)[name = string("input_293_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_291_cast_fp16)[name = string("input_293_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590709184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592806400)))];
-            tensor<fp16, [1, 256, 1024]> linear_96_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_96_cast_fp16")];
-            tensor<int32, [4]> var_2817 = const()[name = string("op_2817"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_243_cast_fp16 = reshape(shape = var_2817, x = linear_96_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_96_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_96_cast_fp16")];
+            tensor<int32, [4]> var_2817 = const()[name = string("op_2817"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_243_cast_fp16 = reshape(shape = var_2817, x = linear_96_cast_fp16)[name = string("x_243_cast_fp16")];
             tensor<int32, [4]> var_2819 = const()[name = string("op_2819"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2824 = const()[name = string("op_2824"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2820_cast_fp16 = transpose(perm = var_2819, x = x_243_cast_fp16)[name = string("transpose_71")];
-            tensor<fp16, [16, 256, 64]> query_layer_75_cast_fp16 = reshape(shape = var_2824, x = var_2820_cast_fp16)[name = string("query_layer_75_cast_fp16")];
+            tensor<int32, [3]> var_2824 = const()[name = string("op_2824"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2820_cast_fp16 = transpose(perm = var_2819, x = x_243_cast_fp16)[name = string("transpose_71")];
+            tensor<fp16, [16, 192, 64]> query_layer_75_cast_fp16 = reshape(shape = var_2824, x = var_2820_cast_fp16)[name = string("query_layer_75_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592808512)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594905728)))];
-            tensor<fp16, [1, 256, 1024]> linear_97_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_97_cast_fp16")];
-            tensor<int32, [4]> var_2831 = const()[name = string("op_2831"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_247_cast_fp16 = reshape(shape = var_2831, x = linear_97_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_97_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<int32, [4]> var_2831 = const()[name = string("op_2831"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_247_cast_fp16 = reshape(shape = var_2831, x = linear_97_cast_fp16)[name = string("x_247_cast_fp16")];
             tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2834_cast_fp16 = transpose(perm = var_2833, x = x_247_cast_fp16)[name = string("transpose_70")];
-            tensor<fp16, [16, 256, 64]> key_layer_25_cast_fp16 = reshape(shape = var_2838, x = var_2834_cast_fp16)[name = string("key_layer_25_cast_fp16")];
+            tensor<int32, [3]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2834_cast_fp16 = transpose(perm = var_2833, x = x_247_cast_fp16)[name = string("transpose_70")];
+            tensor<fp16, [16, 192, 64]> key_layer_25_cast_fp16 = reshape(shape = var_2838, x = var_2834_cast_fp16)[name = string("key_layer_25_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594907840)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597005056)))];
-            tensor<fp16, [1, 256, 1024]> linear_98_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_98_cast_fp16")];
-            tensor<int32, [4]> var_2845 = const()[name = string("op_2845"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_251_cast_fp16 = reshape(shape = var_2845, x = linear_98_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_98_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<int32, [4]> var_2845 = const()[name = string("op_2845"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_251_cast_fp16 = reshape(shape = var_2845, x = linear_98_cast_fp16)[name = string("x_251_cast_fp16")];
             tensor<int32, [4]> var_2847 = const()[name = string("op_2847"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2852 = const()[name = string("op_2852"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2848_cast_fp16 = transpose(perm = var_2847, x = x_251_cast_fp16)[name = string("transpose_69")];
-            tensor<fp16, [16, 256, 64]> value_layer_25_cast_fp16 = reshape(shape = var_2852, x = var_2848_cast_fp16)[name = string("value_layer_25_cast_fp16")];
+            tensor<int32, [3]> var_2852 = const()[name = string("op_2852"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2848_cast_fp16 = transpose(perm = var_2847, x = x_251_cast_fp16)[name = string("transpose_69")];
+            tensor<fp16, [16, 192, 64]> value_layer_25_cast_fp16 = reshape(shape = var_2852, x = var_2848_cast_fp16)[name = string("value_layer_25_cast_fp16")];
             tensor<int32, [3]> var_2862_perm_0 = const()[name = string("op_2862_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2864_y_0_to_fp16 = const()[name = string("_inversed_2864_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_2862_cast_fp16 = transpose(perm = var_2862_perm_0, x = key_layer_25_cast_fp16)[name = string("transpose_68")];
-            tensor<fp16, [16, 64, 256]> _inversed_2864_cast_fp16 = mul(x = var_2862_cast_fp16, y = _inversed_2864_y_0_to_fp16)[name = string("_inversed_2864_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_2862_cast_fp16 = transpose(perm = var_2862_perm_0, x = key_layer_25_cast_fp16)[name = string("transpose_68")];
+            tensor<fp16, [16, 64, 192]> _inversed_2864_cast_fp16 = mul(x = var_2862_cast_fp16, y = _inversed_2864_y_0_to_fp16)[name = string("_inversed_2864_cast_fp16")];
             bool attention_scores_73_transpose_x_0 = const()[name = string("attention_scores_73_transpose_x_0"), val = bool(false)];
             bool attention_scores_73_transpose_y_0 = const()[name = string("attention_scores_73_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_73_cast_fp16 = matmul(transpose_x = attention_scores_73_transpose_x_0, transpose_y = attention_scores_73_transpose_y_0, x = query_layer_75_cast_fp16, y = _inversed_2864_cast_fp16)[name = string("attention_scores_73_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_73_cast_fp16 = matmul(transpose_x = attention_scores_73_transpose_x_0, transpose_y = attention_scores_73_transpose_y_0, x = query_layer_75_cast_fp16, y = _inversed_2864_cast_fp16)[name = string("attention_scores_73_cast_fp16")];
             bool c2p_att_49_transpose_x_0 = const()[name = string("c2p_att_49_transpose_x_0"), val = bool(false)];
             bool c2p_att_49_transpose_y_0 = const()[name = string("c2p_att_49_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2920_to_fp16 = const()[name = string("op_2920_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597007168)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_49_cast_fp16 = matmul(transpose_x = c2p_att_49_transpose_x_0, transpose_y = c2p_att_49_transpose_y_0, x = query_layer_75_cast_fp16, y = var_2920_to_fp16)[name = string("c2p_att_49_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_49_cast_fp16 = matmul(transpose_x = c2p_att_49_transpose_x_0, transpose_y = c2p_att_49_transpose_y_0, x = query_layer_75_cast_fp16, y = var_2920_to_fp16)[name = string("c2p_att_49_cast_fp16")];
             bool c2p_att_51_validate_indices_0 = const()[name = string("c2p_att_51_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_51_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_51_validate_indices_0, x = c2p_att_49_cast_fp16)[name = string("c2p_att_51_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_51_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_51_validate_indices_0, x = c2p_att_49_cast_fp16)[name = string("c2p_att_51_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2933_y_0_to_fp16 = const()[name = string("_inversed_2933_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_2933_cast_fp16 = mul(x = c2p_att_51_cast_fp16_cast_int16, y = _inversed_2933_y_0_to_fp16)[name = string("_inversed_2933_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_2933_cast_fp16 = mul(x = c2p_att_51_cast_fp16_cast_int16, y = _inversed_2933_y_0_to_fp16)[name = string("_inversed_2933_cast_fp16")];
             bool p2c_att_49_transpose_x_0 = const()[name = string("p2c_att_49_transpose_x_0"), val = bool(false)];
             bool p2c_att_49_transpose_y_0 = const()[name = string("p2c_att_49_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2948_to_fp16 = const()[name = string("op_2948_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598055808)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_49_cast_fp16 = matmul(transpose_x = p2c_att_49_transpose_x_0, transpose_y = p2c_att_49_transpose_y_0, x = key_layer_25_cast_fp16, y = var_2948_to_fp16)[name = string("p2c_att_49_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_49_cast_fp16 = matmul(transpose_x = p2c_att_49_transpose_x_0, transpose_y = p2c_att_49_transpose_y_0, x = key_layer_25_cast_fp16, y = var_2948_to_fp16)[name = string("p2c_att_49_cast_fp16")];
             bool var_2956_validate_indices_0 = const()[name = string("op_2956_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_2956_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2956_validate_indices_0, x = p2c_att_49_cast_fp16)[name = string("op_2956_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_2956_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2956_validate_indices_0, x = p2c_att_49_cast_fp16)[name = string("op_2956_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_51_perm_0 = const()[name = string("p2c_att_51_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2959_y_0_to_fp16 = const()[name = string("_inversed_2959_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_51_cast_fp16 = transpose(perm = p2c_att_51_perm_0, x = var_2956_cast_fp16_cast_int16)[name = string("transpose_67")];
-            tensor<fp16, [16, 256, 256]> _inversed_2959_cast_fp16 = mul(x = p2c_att_51_cast_fp16, y = _inversed_2959_y_0_to_fp16)[name = string("_inversed_2959_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_25_cast_fp16 = add(x = _inversed_2933_cast_fp16, y = _inversed_2959_cast_fp16)[name = string("rel_att_25_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_75_cast_fp16 = add(x = attention_scores_73_cast_fp16, y = rel_att_25_cast_fp16)[name = string("attention_scores_75_cast_fp16")];
-            tensor<int32, [4]> var_2964 = const()[name = string("op_2964"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_77_cast_fp16 = reshape(shape = var_2964, x = attention_scores_75_cast_fp16)[name = string("attention_scores_77_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_297_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_77_cast_fp16, cond = var_339)[name = string("input_297_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_299_cast_fp16 = softmax(axis = var_56, x = input_297_cast_fp16)[name = string("input_299_cast_fp16")];
-            tensor<int32, [3]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_2974_cast_fp16 = reshape(shape = var_2973, x = input_299_cast_fp16)[name = string("op_2974_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_51_cast_fp16 = transpose(perm = p2c_att_51_perm_0, x = var_2956_cast_fp16_cast_int16)[name = string("transpose_67")];
+            tensor<fp16, [16, 192, 192]> _inversed_2959_cast_fp16 = mul(x = p2c_att_51_cast_fp16, y = _inversed_2959_y_0_to_fp16)[name = string("_inversed_2959_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_25_cast_fp16 = add(x = _inversed_2933_cast_fp16, y = _inversed_2959_cast_fp16)[name = string("rel_att_25_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_75_cast_fp16 = add(x = attention_scores_73_cast_fp16, y = rel_att_25_cast_fp16)[name = string("attention_scores_75_cast_fp16")];
+            tensor<int32, [4]> var_2964 = const()[name = string("op_2964"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_77_cast_fp16 = reshape(shape = var_2964, x = attention_scores_75_cast_fp16)[name = string("attention_scores_77_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_297_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_77_cast_fp16, cond = var_339)[name = string("input_297_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_299_cast_fp16 = softmax(axis = var_56, x = input_297_cast_fp16)[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_2974_cast_fp16 = reshape(shape = var_2973, x = input_299_cast_fp16)[name = string("op_2974_cast_fp16")];
             bool context_layer_49_transpose_x_0 = const()[name = string("context_layer_49_transpose_x_0"), val = bool(false)];
             bool context_layer_49_transpose_y_0 = const()[name = string("context_layer_49_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_49_cast_fp16 = matmul(transpose_x = context_layer_49_transpose_x_0, transpose_y = context_layer_49_transpose_y_0, x = var_2974_cast_fp16, y = value_layer_25_cast_fp16)[name = string("context_layer_49_cast_fp16")];
-            tensor<int32, [4]> var_2978 = const()[name = string("op_2978"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_2979_cast_fp16 = reshape(shape = var_2978, x = context_layer_49_cast_fp16)[name = string("op_2979_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_49_cast_fp16 = matmul(transpose_x = context_layer_49_transpose_x_0, transpose_y = context_layer_49_transpose_y_0, x = var_2974_cast_fp16, y = value_layer_25_cast_fp16)[name = string("context_layer_49_cast_fp16")];
+            tensor<int32, [4]> var_2978 = const()[name = string("op_2978"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_2979_cast_fp16 = reshape(shape = var_2978, x = context_layer_49_cast_fp16)[name = string("op_2979_cast_fp16")];
             tensor<int32, [4]> var_2980 = const()[name = string("op_2980"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2985 = const()[name = string("op_2985"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_2981_cast_fp16 = transpose(perm = var_2980, x = var_2979_cast_fp16)[name = string("transpose_66")];
-            tensor<fp16, [1, 256, 1024]> input_301_cast_fp16 = reshape(shape = var_2985, x = var_2981_cast_fp16)[name = string("input_301_cast_fp16")];
+            tensor<int32, [3]> var_2985 = const()[name = string("op_2985"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_2981_cast_fp16 = transpose(perm = var_2980, x = var_2979_cast_fp16)[name = string("transpose_66")];
+            tensor<fp16, [1, 192, 1024]> input_301_cast_fp16 = reshape(shape = var_2985, x = var_2981_cast_fp16)[name = string("input_301_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599104448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601201664)))];
-            tensor<fp16, [1, 256, 1024]> linear_101_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_attention_output_dense_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_101_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_305_cast_fp16 = add(x = linear_101_cast_fp16, y = input_293_cast_fp16)[name = string("input_305_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_101_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_attention_output_dense_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_101_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_305_cast_fp16 = add(x = linear_101_cast_fp16, y = input_293_cast_fp16)[name = string("input_305_cast_fp16")];
             tensor<int32, [1]> input_307_axes_0 = const()[name = string("input_307_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601203776)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601205888)))];
-            tensor<fp16, [1, 256, 1024]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_12_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601208000)))];
             tensor<fp16, [4096]> encoder_encoder_layer_12_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609596672)))];
-            tensor<fp16, [1, 256, 4096]> linear_102_cast_fp16 = linear(bias = encoder_encoder_layer_12_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_12_intermediate_dense_weight_to_fp16, x = input_307_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_102_cast_fp16 = linear(bias = encoder_encoder_layer_12_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_12_intermediate_dense_weight_to_fp16, x = input_307_cast_fp16)[name = string("linear_102_cast_fp16")];
             string input_311_mode_0 = const()[name = string("input_311_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = linear_102_cast_fp16)[name = string("input_311_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = linear_102_cast_fp16)[name = string("input_311_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_12_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609604928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617993600)))];
-            tensor<fp16, [1, 256, 1024]> linear_103_cast_fp16 = linear(bias = encoder_encoder_layer_12_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_output_dense_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_103_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_315_cast_fp16 = add(x = linear_103_cast_fp16, y = input_307_cast_fp16)[name = string("input_315_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_103_cast_fp16 = linear(bias = encoder_encoder_layer_12_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_output_dense_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_315_cast_fp16 = add(x = linear_103_cast_fp16, y = input_307_cast_fp16)[name = string("input_315_cast_fp16")];
             tensor<int32, [1]> input_317_axes_0 = const()[name = string("input_317_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617995712)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617997824)))];
-            tensor<fp16, [1, 256, 1024]> input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16, x = input_315_cast_fp16)[name = string("input_317_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16, x = input_315_cast_fp16)[name = string("input_317_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617999936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620097152)))];
-            tensor<fp16, [1, 256, 1024]> linear_104_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_104_cast_fp16")];
-            tensor<int32, [4]> var_3036 = const()[name = string("op_3036"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_263_cast_fp16 = reshape(shape = var_3036, x = linear_104_cast_fp16)[name = string("x_263_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_104_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_3036 = const()[name = string("op_3036"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_263_cast_fp16 = reshape(shape = var_3036, x = linear_104_cast_fp16)[name = string("x_263_cast_fp16")];
             tensor<int32, [4]> var_3038 = const()[name = string("op_3038"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3043 = const()[name = string("op_3043"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3039_cast_fp16 = transpose(perm = var_3038, x = x_263_cast_fp16)[name = string("transpose_65")];
-            tensor<fp16, [16, 256, 64]> query_layer_81_cast_fp16 = reshape(shape = var_3043, x = var_3039_cast_fp16)[name = string("query_layer_81_cast_fp16")];
+            tensor<int32, [3]> var_3043 = const()[name = string("op_3043"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3039_cast_fp16 = transpose(perm = var_3038, x = x_263_cast_fp16)[name = string("transpose_65")];
+            tensor<fp16, [16, 192, 64]> query_layer_81_cast_fp16 = reshape(shape = var_3043, x = var_3039_cast_fp16)[name = string("query_layer_81_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620099264)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622196480)))];
-            tensor<fp16, [1, 256, 1024]> linear_105_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_105_cast_fp16")];
-            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_267_cast_fp16 = reshape(shape = var_3050, x = linear_105_cast_fp16)[name = string("x_267_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_105_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_105_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_267_cast_fp16 = reshape(shape = var_3050, x = linear_105_cast_fp16)[name = string("x_267_cast_fp16")];
             tensor<int32, [4]> var_3052 = const()[name = string("op_3052"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3053_cast_fp16 = transpose(perm = var_3052, x = x_267_cast_fp16)[name = string("transpose_64")];
-            tensor<fp16, [16, 256, 64]> key_layer_27_cast_fp16 = reshape(shape = var_3057, x = var_3053_cast_fp16)[name = string("key_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3053_cast_fp16 = transpose(perm = var_3052, x = x_267_cast_fp16)[name = string("transpose_64")];
+            tensor<fp16, [16, 192, 64]> key_layer_27_cast_fp16 = reshape(shape = var_3057, x = var_3053_cast_fp16)[name = string("key_layer_27_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622198592)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624295808)))];
-            tensor<fp16, [1, 256, 1024]> linear_106_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_106_cast_fp16")];
-            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_271_cast_fp16 = reshape(shape = var_3064, x = linear_106_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_106_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_271_cast_fp16 = reshape(shape = var_3064, x = linear_106_cast_fp16)[name = string("x_271_cast_fp16")];
             tensor<int32, [4]> var_3066 = const()[name = string("op_3066"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3071 = const()[name = string("op_3071"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3067_cast_fp16 = transpose(perm = var_3066, x = x_271_cast_fp16)[name = string("transpose_63")];
-            tensor<fp16, [16, 256, 64]> value_layer_27_cast_fp16 = reshape(shape = var_3071, x = var_3067_cast_fp16)[name = string("value_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_3071 = const()[name = string("op_3071"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3067_cast_fp16 = transpose(perm = var_3066, x = x_271_cast_fp16)[name = string("transpose_63")];
+            tensor<fp16, [16, 192, 64]> value_layer_27_cast_fp16 = reshape(shape = var_3071, x = var_3067_cast_fp16)[name = string("value_layer_27_cast_fp16")];
             tensor<int32, [3]> var_3081_perm_0 = const()[name = string("op_3081_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3083_y_0_to_fp16 = const()[name = string("_inversed_3083_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_3081_cast_fp16 = transpose(perm = var_3081_perm_0, x = key_layer_27_cast_fp16)[name = string("transpose_62")];
-            tensor<fp16, [16, 64, 256]> _inversed_3083_cast_fp16 = mul(x = var_3081_cast_fp16, y = _inversed_3083_y_0_to_fp16)[name = string("_inversed_3083_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_3081_cast_fp16 = transpose(perm = var_3081_perm_0, x = key_layer_27_cast_fp16)[name = string("transpose_62")];
+            tensor<fp16, [16, 64, 192]> _inversed_3083_cast_fp16 = mul(x = var_3081_cast_fp16, y = _inversed_3083_y_0_to_fp16)[name = string("_inversed_3083_cast_fp16")];
             bool attention_scores_79_transpose_x_0 = const()[name = string("attention_scores_79_transpose_x_0"), val = bool(false)];
             bool attention_scores_79_transpose_y_0 = const()[name = string("attention_scores_79_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_79_cast_fp16 = matmul(transpose_x = attention_scores_79_transpose_x_0, transpose_y = attention_scores_79_transpose_y_0, x = query_layer_81_cast_fp16, y = _inversed_3083_cast_fp16)[name = string("attention_scores_79_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_79_cast_fp16 = matmul(transpose_x = attention_scores_79_transpose_x_0, transpose_y = attention_scores_79_transpose_y_0, x = query_layer_81_cast_fp16, y = _inversed_3083_cast_fp16)[name = string("attention_scores_79_cast_fp16")];
             bool c2p_att_53_transpose_x_0 = const()[name = string("c2p_att_53_transpose_x_0"), val = bool(false)];
             bool c2p_att_53_transpose_y_0 = const()[name = string("c2p_att_53_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3139_to_fp16 = const()[name = string("op_3139_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624297920)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_53_cast_fp16 = matmul(transpose_x = c2p_att_53_transpose_x_0, transpose_y = c2p_att_53_transpose_y_0, x = query_layer_81_cast_fp16, y = var_3139_to_fp16)[name = string("c2p_att_53_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_53_cast_fp16 = matmul(transpose_x = c2p_att_53_transpose_x_0, transpose_y = c2p_att_53_transpose_y_0, x = query_layer_81_cast_fp16, y = var_3139_to_fp16)[name = string("c2p_att_53_cast_fp16")];
             bool c2p_att_55_validate_indices_0 = const()[name = string("c2p_att_55_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_55_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_55_validate_indices_0, x = c2p_att_53_cast_fp16)[name = string("c2p_att_55_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_55_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_55_validate_indices_0, x = c2p_att_53_cast_fp16)[name = string("c2p_att_55_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3152_y_0_to_fp16 = const()[name = string("_inversed_3152_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_3152_cast_fp16 = mul(x = c2p_att_55_cast_fp16_cast_int16, y = _inversed_3152_y_0_to_fp16)[name = string("_inversed_3152_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_3152_cast_fp16 = mul(x = c2p_att_55_cast_fp16_cast_int16, y = _inversed_3152_y_0_to_fp16)[name = string("_inversed_3152_cast_fp16")];
             bool p2c_att_53_transpose_x_0 = const()[name = string("p2c_att_53_transpose_x_0"), val = bool(false)];
             bool p2c_att_53_transpose_y_0 = const()[name = string("p2c_att_53_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3167_to_fp16 = const()[name = string("op_3167_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(625346560)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_53_cast_fp16 = matmul(transpose_x = p2c_att_53_transpose_x_0, transpose_y = p2c_att_53_transpose_y_0, x = key_layer_27_cast_fp16, y = var_3167_to_fp16)[name = string("p2c_att_53_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_53_cast_fp16 = matmul(transpose_x = p2c_att_53_transpose_x_0, transpose_y = p2c_att_53_transpose_y_0, x = key_layer_27_cast_fp16, y = var_3167_to_fp16)[name = string("p2c_att_53_cast_fp16")];
             bool var_3175_validate_indices_0 = const()[name = string("op_3175_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_3175_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3175_validate_indices_0, x = p2c_att_53_cast_fp16)[name = string("op_3175_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_3175_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3175_validate_indices_0, x = p2c_att_53_cast_fp16)[name = string("op_3175_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_55_perm_0 = const()[name = string("p2c_att_55_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3178_y_0_to_fp16 = const()[name = string("_inversed_3178_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_55_cast_fp16 = transpose(perm = p2c_att_55_perm_0, x = var_3175_cast_fp16_cast_int16)[name = string("transpose_61")];
-            tensor<fp16, [16, 256, 256]> _inversed_3178_cast_fp16 = mul(x = p2c_att_55_cast_fp16, y = _inversed_3178_y_0_to_fp16)[name = string("_inversed_3178_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_27_cast_fp16 = add(x = _inversed_3152_cast_fp16, y = _inversed_3178_cast_fp16)[name = string("rel_att_27_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_81_cast_fp16 = add(x = attention_scores_79_cast_fp16, y = rel_att_27_cast_fp16)[name = string("attention_scores_81_cast_fp16")];
-            tensor<int32, [4]> var_3183 = const()[name = string("op_3183"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_83_cast_fp16 = reshape(shape = var_3183, x = attention_scores_81_cast_fp16)[name = string("attention_scores_83_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_321_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_83_cast_fp16, cond = var_339)[name = string("input_321_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_323_cast_fp16 = softmax(axis = var_56, x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
-            tensor<int32, [3]> var_3192 = const()[name = string("op_3192"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_3193_cast_fp16 = reshape(shape = var_3192, x = input_323_cast_fp16)[name = string("op_3193_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_55_cast_fp16 = transpose(perm = p2c_att_55_perm_0, x = var_3175_cast_fp16_cast_int16)[name = string("transpose_61")];
+            tensor<fp16, [16, 192, 192]> _inversed_3178_cast_fp16 = mul(x = p2c_att_55_cast_fp16, y = _inversed_3178_y_0_to_fp16)[name = string("_inversed_3178_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_27_cast_fp16 = add(x = _inversed_3152_cast_fp16, y = _inversed_3178_cast_fp16)[name = string("rel_att_27_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_81_cast_fp16 = add(x = attention_scores_79_cast_fp16, y = rel_att_27_cast_fp16)[name = string("attention_scores_81_cast_fp16")];
+            tensor<int32, [4]> var_3183 = const()[name = string("op_3183"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_83_cast_fp16 = reshape(shape = var_3183, x = attention_scores_81_cast_fp16)[name = string("attention_scores_83_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_321_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_83_cast_fp16, cond = var_339)[name = string("input_321_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_323_cast_fp16 = softmax(axis = var_56, x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [3]> var_3192 = const()[name = string("op_3192"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_3193_cast_fp16 = reshape(shape = var_3192, x = input_323_cast_fp16)[name = string("op_3193_cast_fp16")];
             bool context_layer_53_transpose_x_0 = const()[name = string("context_layer_53_transpose_x_0"), val = bool(false)];
             bool context_layer_53_transpose_y_0 = const()[name = string("context_layer_53_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_53_cast_fp16 = matmul(transpose_x = context_layer_53_transpose_x_0, transpose_y = context_layer_53_transpose_y_0, x = var_3193_cast_fp16, y = value_layer_27_cast_fp16)[name = string("context_layer_53_cast_fp16")];
-            tensor<int32, [4]> var_3197 = const()[name = string("op_3197"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3198_cast_fp16 = reshape(shape = var_3197, x = context_layer_53_cast_fp16)[name = string("op_3198_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_53_cast_fp16 = matmul(transpose_x = context_layer_53_transpose_x_0, transpose_y = context_layer_53_transpose_y_0, x = var_3193_cast_fp16, y = value_layer_27_cast_fp16)[name = string("context_layer_53_cast_fp16")];
+            tensor<int32, [4]> var_3197 = const()[name = string("op_3197"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3198_cast_fp16 = reshape(shape = var_3197, x = context_layer_53_cast_fp16)[name = string("op_3198_cast_fp16")];
             tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3204 = const()[name = string("op_3204"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_3200_cast_fp16 = transpose(perm = var_3199, x = var_3198_cast_fp16)[name = string("transpose_60")];
-            tensor<fp16, [1, 256, 1024]> input_325_cast_fp16 = reshape(shape = var_3204, x = var_3200_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<int32, [3]> var_3204 = const()[name = string("op_3204"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_3200_cast_fp16 = transpose(perm = var_3199, x = var_3198_cast_fp16)[name = string("transpose_60")];
+            tensor<fp16, [1, 192, 1024]> input_325_cast_fp16 = reshape(shape = var_3204, x = var_3200_cast_fp16)[name = string("input_325_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626395200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628492416)))];
-            tensor<fp16, [1, 256, 1024]> linear_109_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_attention_output_dense_weight_to_fp16, x = input_325_cast_fp16)[name = string("linear_109_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_329_cast_fp16 = add(x = linear_109_cast_fp16, y = input_317_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_109_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_attention_output_dense_weight_to_fp16, x = input_325_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_329_cast_fp16 = add(x = linear_109_cast_fp16, y = input_317_cast_fp16)[name = string("input_329_cast_fp16")];
             tensor<int32, [1]> input_331_axes_0 = const()[name = string("input_331_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628494528)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628496640)))];
-            tensor<fp16, [1, 256, 1024]> input_331_cast_fp16 = layer_norm(axes = input_331_axes_0, beta = encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16, x = input_329_cast_fp16)[name = string("input_331_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_331_cast_fp16 = layer_norm(axes = input_331_axes_0, beta = encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16, x = input_329_cast_fp16)[name = string("input_331_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_13_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628498752)))];
             tensor<fp16, [4096]> encoder_encoder_layer_13_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636887424)))];
-            tensor<fp16, [1, 256, 4096]> linear_110_cast_fp16 = linear(bias = encoder_encoder_layer_13_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_13_intermediate_dense_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_110_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_110_cast_fp16 = linear(bias = encoder_encoder_layer_13_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_13_intermediate_dense_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_110_cast_fp16")];
             string input_335_mode_0 = const()[name = string("input_335_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_335_cast_fp16 = gelu(mode = input_335_mode_0, x = linear_110_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_335_cast_fp16 = gelu(mode = input_335_mode_0, x = linear_110_cast_fp16)[name = string("input_335_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_13_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636895680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645284352)))];
-            tensor<fp16, [1, 256, 1024]> linear_111_cast_fp16 = linear(bias = encoder_encoder_layer_13_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_output_dense_weight_to_fp16, x = input_335_cast_fp16)[name = string("linear_111_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_339_cast_fp16 = add(x = linear_111_cast_fp16, y = input_331_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_111_cast_fp16 = linear(bias = encoder_encoder_layer_13_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_output_dense_weight_to_fp16, x = input_335_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_339_cast_fp16 = add(x = linear_111_cast_fp16, y = input_331_cast_fp16)[name = string("input_339_cast_fp16")];
             tensor<int32, [1]> input_341_axes_0 = const()[name = string("input_341_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645286464)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645288576)))];
-            tensor<fp16, [1, 256, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_0, beta = encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_0, beta = encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645290688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647387904)))];
-            tensor<fp16, [1, 256, 1024]> linear_112_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_112_cast_fp16")];
-            tensor<int32, [4]> var_3255 = const()[name = string("op_3255"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_283_cast_fp16 = reshape(shape = var_3255, x = linear_112_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_112_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3255 = const()[name = string("op_3255"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_283_cast_fp16 = reshape(shape = var_3255, x = linear_112_cast_fp16)[name = string("x_283_cast_fp16")];
             tensor<int32, [4]> var_3257 = const()[name = string("op_3257"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3262 = const()[name = string("op_3262"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3258_cast_fp16 = transpose(perm = var_3257, x = x_283_cast_fp16)[name = string("transpose_59")];
-            tensor<fp16, [16, 256, 64]> query_layer_87_cast_fp16 = reshape(shape = var_3262, x = var_3258_cast_fp16)[name = string("query_layer_87_cast_fp16")];
+            tensor<int32, [3]> var_3262 = const()[name = string("op_3262"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3258_cast_fp16 = transpose(perm = var_3257, x = x_283_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [16, 192, 64]> query_layer_87_cast_fp16 = reshape(shape = var_3262, x = var_3258_cast_fp16)[name = string("query_layer_87_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647390016)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(649487232)))];
-            tensor<fp16, [1, 256, 1024]> linear_113_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_113_cast_fp16")];
-            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_287_cast_fp16 = reshape(shape = var_3269, x = linear_113_cast_fp16)[name = string("x_287_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_113_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_287_cast_fp16 = reshape(shape = var_3269, x = linear_113_cast_fp16)[name = string("x_287_cast_fp16")];
             tensor<int32, [4]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3276 = const()[name = string("op_3276"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3272_cast_fp16 = transpose(perm = var_3271, x = x_287_cast_fp16)[name = string("transpose_58")];
-            tensor<fp16, [16, 256, 64]> key_layer_29_cast_fp16 = reshape(shape = var_3276, x = var_3272_cast_fp16)[name = string("key_layer_29_cast_fp16")];
+            tensor<int32, [3]> var_3276 = const()[name = string("op_3276"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3272_cast_fp16 = transpose(perm = var_3271, x = x_287_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [16, 192, 64]> key_layer_29_cast_fp16 = reshape(shape = var_3276, x = var_3272_cast_fp16)[name = string("key_layer_29_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(649489344)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651586560)))];
-            tensor<fp16, [1, 256, 1024]> linear_114_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_114_cast_fp16")];
-            tensor<int32, [4]> var_3283 = const()[name = string("op_3283"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_291_cast_fp16 = reshape(shape = var_3283, x = linear_114_cast_fp16)[name = string("x_291_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_114_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_114_cast_fp16")];
+            tensor<int32, [4]> var_3283 = const()[name = string("op_3283"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_291_cast_fp16 = reshape(shape = var_3283, x = linear_114_cast_fp16)[name = string("x_291_cast_fp16")];
             tensor<int32, [4]> var_3285 = const()[name = string("op_3285"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3286_cast_fp16 = transpose(perm = var_3285, x = x_291_cast_fp16)[name = string("transpose_57")];
-            tensor<fp16, [16, 256, 64]> value_layer_29_cast_fp16 = reshape(shape = var_3290, x = var_3286_cast_fp16)[name = string("value_layer_29_cast_fp16")];
+            tensor<int32, [3]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3286_cast_fp16 = transpose(perm = var_3285, x = x_291_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [16, 192, 64]> value_layer_29_cast_fp16 = reshape(shape = var_3290, x = var_3286_cast_fp16)[name = string("value_layer_29_cast_fp16")];
             tensor<int32, [3]> var_3300_perm_0 = const()[name = string("op_3300_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3302_y_0_to_fp16 = const()[name = string("_inversed_3302_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_3300_cast_fp16 = transpose(perm = var_3300_perm_0, x = key_layer_29_cast_fp16)[name = string("transpose_56")];
-            tensor<fp16, [16, 64, 256]> _inversed_3302_cast_fp16 = mul(x = var_3300_cast_fp16, y = _inversed_3302_y_0_to_fp16)[name = string("_inversed_3302_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_3300_cast_fp16 = transpose(perm = var_3300_perm_0, x = key_layer_29_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [16, 64, 192]> _inversed_3302_cast_fp16 = mul(x = var_3300_cast_fp16, y = _inversed_3302_y_0_to_fp16)[name = string("_inversed_3302_cast_fp16")];
             bool attention_scores_85_transpose_x_0 = const()[name = string("attention_scores_85_transpose_x_0"), val = bool(false)];
             bool attention_scores_85_transpose_y_0 = const()[name = string("attention_scores_85_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_85_cast_fp16 = matmul(transpose_x = attention_scores_85_transpose_x_0, transpose_y = attention_scores_85_transpose_y_0, x = query_layer_87_cast_fp16, y = _inversed_3302_cast_fp16)[name = string("attention_scores_85_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_85_cast_fp16 = matmul(transpose_x = attention_scores_85_transpose_x_0, transpose_y = attention_scores_85_transpose_y_0, x = query_layer_87_cast_fp16, y = _inversed_3302_cast_fp16)[name = string("attention_scores_85_cast_fp16")];
             bool c2p_att_57_transpose_x_0 = const()[name = string("c2p_att_57_transpose_x_0"), val = bool(false)];
             bool c2p_att_57_transpose_y_0 = const()[name = string("c2p_att_57_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3358_to_fp16 = const()[name = string("op_3358_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651588672)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_57_cast_fp16 = matmul(transpose_x = c2p_att_57_transpose_x_0, transpose_y = c2p_att_57_transpose_y_0, x = query_layer_87_cast_fp16, y = var_3358_to_fp16)[name = string("c2p_att_57_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_57_cast_fp16 = matmul(transpose_x = c2p_att_57_transpose_x_0, transpose_y = c2p_att_57_transpose_y_0, x = query_layer_87_cast_fp16, y = var_3358_to_fp16)[name = string("c2p_att_57_cast_fp16")];
             bool c2p_att_59_validate_indices_0 = const()[name = string("c2p_att_59_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_59_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_59_validate_indices_0, x = c2p_att_57_cast_fp16)[name = string("c2p_att_59_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_59_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_59_validate_indices_0, x = c2p_att_57_cast_fp16)[name = string("c2p_att_59_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3371_y_0_to_fp16 = const()[name = string("_inversed_3371_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_3371_cast_fp16 = mul(x = c2p_att_59_cast_fp16_cast_int16, y = _inversed_3371_y_0_to_fp16)[name = string("_inversed_3371_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_3371_cast_fp16 = mul(x = c2p_att_59_cast_fp16_cast_int16, y = _inversed_3371_y_0_to_fp16)[name = string("_inversed_3371_cast_fp16")];
             bool p2c_att_57_transpose_x_0 = const()[name = string("p2c_att_57_transpose_x_0"), val = bool(false)];
             bool p2c_att_57_transpose_y_0 = const()[name = string("p2c_att_57_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3386_to_fp16 = const()[name = string("op_3386_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652637312)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_57_cast_fp16 = matmul(transpose_x = p2c_att_57_transpose_x_0, transpose_y = p2c_att_57_transpose_y_0, x = key_layer_29_cast_fp16, y = var_3386_to_fp16)[name = string("p2c_att_57_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_57_cast_fp16 = matmul(transpose_x = p2c_att_57_transpose_x_0, transpose_y = p2c_att_57_transpose_y_0, x = key_layer_29_cast_fp16, y = var_3386_to_fp16)[name = string("p2c_att_57_cast_fp16")];
             bool var_3394_validate_indices_0 = const()[name = string("op_3394_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_3394_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3394_validate_indices_0, x = p2c_att_57_cast_fp16)[name = string("op_3394_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_3394_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3394_validate_indices_0, x = p2c_att_57_cast_fp16)[name = string("op_3394_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_59_perm_0 = const()[name = string("p2c_att_59_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3397_y_0_to_fp16 = const()[name = string("_inversed_3397_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_59_cast_fp16 = transpose(perm = p2c_att_59_perm_0, x = var_3394_cast_fp16_cast_int16)[name = string("transpose_55")];
-            tensor<fp16, [16, 256, 256]> _inversed_3397_cast_fp16 = mul(x = p2c_att_59_cast_fp16, y = _inversed_3397_y_0_to_fp16)[name = string("_inversed_3397_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_29_cast_fp16 = add(x = _inversed_3371_cast_fp16, y = _inversed_3397_cast_fp16)[name = string("rel_att_29_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_87_cast_fp16 = add(x = attention_scores_85_cast_fp16, y = rel_att_29_cast_fp16)[name = string("attention_scores_87_cast_fp16")];
-            tensor<int32, [4]> var_3402 = const()[name = string("op_3402"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_89_cast_fp16 = reshape(shape = var_3402, x = attention_scores_87_cast_fp16)[name = string("attention_scores_89_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_345_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_89_cast_fp16, cond = var_339)[name = string("input_345_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_347_cast_fp16 = softmax(axis = var_56, x = input_345_cast_fp16)[name = string("input_347_cast_fp16")];
-            tensor<int32, [3]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_3412_cast_fp16 = reshape(shape = var_3411, x = input_347_cast_fp16)[name = string("op_3412_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_59_cast_fp16 = transpose(perm = p2c_att_59_perm_0, x = var_3394_cast_fp16_cast_int16)[name = string("transpose_55")];
+            tensor<fp16, [16, 192, 192]> _inversed_3397_cast_fp16 = mul(x = p2c_att_59_cast_fp16, y = _inversed_3397_y_0_to_fp16)[name = string("_inversed_3397_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_29_cast_fp16 = add(x = _inversed_3371_cast_fp16, y = _inversed_3397_cast_fp16)[name = string("rel_att_29_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_87_cast_fp16 = add(x = attention_scores_85_cast_fp16, y = rel_att_29_cast_fp16)[name = string("attention_scores_87_cast_fp16")];
+            tensor<int32, [4]> var_3402 = const()[name = string("op_3402"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_89_cast_fp16 = reshape(shape = var_3402, x = attention_scores_87_cast_fp16)[name = string("attention_scores_89_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_345_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_89_cast_fp16, cond = var_339)[name = string("input_345_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_347_cast_fp16 = softmax(axis = var_56, x = input_345_cast_fp16)[name = string("input_347_cast_fp16")];
+            tensor<int32, [3]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_3412_cast_fp16 = reshape(shape = var_3411, x = input_347_cast_fp16)[name = string("op_3412_cast_fp16")];
             bool context_layer_57_transpose_x_0 = const()[name = string("context_layer_57_transpose_x_0"), val = bool(false)];
             bool context_layer_57_transpose_y_0 = const()[name = string("context_layer_57_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_57_cast_fp16 = matmul(transpose_x = context_layer_57_transpose_x_0, transpose_y = context_layer_57_transpose_y_0, x = var_3412_cast_fp16, y = value_layer_29_cast_fp16)[name = string("context_layer_57_cast_fp16")];
-            tensor<int32, [4]> var_3416 = const()[name = string("op_3416"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3417_cast_fp16 = reshape(shape = var_3416, x = context_layer_57_cast_fp16)[name = string("op_3417_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_57_cast_fp16 = matmul(transpose_x = context_layer_57_transpose_x_0, transpose_y = context_layer_57_transpose_y_0, x = var_3412_cast_fp16, y = value_layer_29_cast_fp16)[name = string("context_layer_57_cast_fp16")];
+            tensor<int32, [4]> var_3416 = const()[name = string("op_3416"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3417_cast_fp16 = reshape(shape = var_3416, x = context_layer_57_cast_fp16)[name = string("op_3417_cast_fp16")];
             tensor<int32, [4]> var_3418 = const()[name = string("op_3418"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3423 = const()[name = string("op_3423"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_3419_cast_fp16 = transpose(perm = var_3418, x = var_3417_cast_fp16)[name = string("transpose_54")];
-            tensor<fp16, [1, 256, 1024]> input_349_cast_fp16 = reshape(shape = var_3423, x = var_3419_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [3]> var_3423 = const()[name = string("op_3423"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_3419_cast_fp16 = transpose(perm = var_3418, x = var_3417_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 192, 1024]> input_349_cast_fp16 = reshape(shape = var_3423, x = var_3419_cast_fp16)[name = string("input_349_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(653685952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655783168)))];
-            tensor<fp16, [1, 256, 1024]> linear_117_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_attention_output_dense_weight_to_fp16, x = input_349_cast_fp16)[name = string("linear_117_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_353_cast_fp16 = add(x = linear_117_cast_fp16, y = input_341_cast_fp16)[name = string("input_353_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_117_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_attention_output_dense_weight_to_fp16, x = input_349_cast_fp16)[name = string("linear_117_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_353_cast_fp16 = add(x = linear_117_cast_fp16, y = input_341_cast_fp16)[name = string("input_353_cast_fp16")];
             tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655785280)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655787392)))];
-            tensor<fp16, [1, 256, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, beta = encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, beta = encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_14_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655789504)))];
             tensor<fp16, [4096]> encoder_encoder_layer_14_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664178176)))];
-            tensor<fp16, [1, 256, 4096]> linear_118_cast_fp16 = linear(bias = encoder_encoder_layer_14_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_14_intermediate_dense_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_118_cast_fp16 = linear(bias = encoder_encoder_layer_14_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_14_intermediate_dense_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_118_cast_fp16")];
             string input_359_mode_0 = const()[name = string("input_359_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_359_cast_fp16 = gelu(mode = input_359_mode_0, x = linear_118_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_359_cast_fp16 = gelu(mode = input_359_mode_0, x = linear_118_cast_fp16)[name = string("input_359_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_14_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664186432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672575104)))];
-            tensor<fp16, [1, 256, 1024]> linear_119_cast_fp16 = linear(bias = encoder_encoder_layer_14_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_output_dense_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_119_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_363_cast_fp16 = add(x = linear_119_cast_fp16, y = input_355_cast_fp16)[name = string("input_363_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_119_cast_fp16 = linear(bias = encoder_encoder_layer_14_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_output_dense_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_119_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_363_cast_fp16 = add(x = linear_119_cast_fp16, y = input_355_cast_fp16)[name = string("input_363_cast_fp16")];
             tensor<int32, [1]> input_365_axes_0 = const()[name = string("input_365_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672577216)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672579328)))];
-            tensor<fp16, [1, 256, 1024]> input_365_cast_fp16 = layer_norm(axes = input_365_axes_0, beta = encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16, x = input_363_cast_fp16)[name = string("input_365_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_365_cast_fp16 = layer_norm(axes = input_365_axes_0, beta = encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16, x = input_363_cast_fp16)[name = string("input_365_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672581440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(674678656)))];
-            tensor<fp16, [1, 256, 1024]> linear_120_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_120_cast_fp16")];
-            tensor<int32, [4]> var_3474 = const()[name = string("op_3474"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_303_cast_fp16 = reshape(shape = var_3474, x = linear_120_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_120_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3474 = const()[name = string("op_3474"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_303_cast_fp16 = reshape(shape = var_3474, x = linear_120_cast_fp16)[name = string("x_303_cast_fp16")];
             tensor<int32, [4]> var_3476 = const()[name = string("op_3476"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3477_cast_fp16 = transpose(perm = var_3476, x = x_303_cast_fp16)[name = string("transpose_53")];
-            tensor<fp16, [16, 256, 64]> query_layer_93_cast_fp16 = reshape(shape = var_3481, x = var_3477_cast_fp16)[name = string("query_layer_93_cast_fp16")];
+            tensor<int32, [3]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3477_cast_fp16 = transpose(perm = var_3476, x = x_303_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [16, 192, 64]> query_layer_93_cast_fp16 = reshape(shape = var_3481, x = var_3477_cast_fp16)[name = string("query_layer_93_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(674680768)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676777984)))];
-            tensor<fp16, [1, 256, 1024]> linear_121_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_121_cast_fp16")];
-            tensor<int32, [4]> var_3488 = const()[name = string("op_3488"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_307_cast_fp16 = reshape(shape = var_3488, x = linear_121_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_121_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3488 = const()[name = string("op_3488"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_307_cast_fp16 = reshape(shape = var_3488, x = linear_121_cast_fp16)[name = string("x_307_cast_fp16")];
             tensor<int32, [4]> var_3490 = const()[name = string("op_3490"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3495 = const()[name = string("op_3495"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3491_cast_fp16 = transpose(perm = var_3490, x = x_307_cast_fp16)[name = string("transpose_52")];
-            tensor<fp16, [16, 256, 64]> key_layer_31_cast_fp16 = reshape(shape = var_3495, x = var_3491_cast_fp16)[name = string("key_layer_31_cast_fp16")];
+            tensor<int32, [3]> var_3495 = const()[name = string("op_3495"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3491_cast_fp16 = transpose(perm = var_3490, x = x_307_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [16, 192, 64]> key_layer_31_cast_fp16 = reshape(shape = var_3495, x = var_3491_cast_fp16)[name = string("key_layer_31_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676780096)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(678877312)))];
-            tensor<fp16, [1, 256, 1024]> linear_122_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_122_cast_fp16")];
-            tensor<int32, [4]> var_3502 = const()[name = string("op_3502"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_311_cast_fp16 = reshape(shape = var_3502, x = linear_122_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_122_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3502 = const()[name = string("op_3502"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_311_cast_fp16 = reshape(shape = var_3502, x = linear_122_cast_fp16)[name = string("x_311_cast_fp16")];
             tensor<int32, [4]> var_3504 = const()[name = string("op_3504"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3509 = const()[name = string("op_3509"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3505_cast_fp16 = transpose(perm = var_3504, x = x_311_cast_fp16)[name = string("transpose_51")];
-            tensor<fp16, [16, 256, 64]> value_layer_31_cast_fp16 = reshape(shape = var_3509, x = var_3505_cast_fp16)[name = string("value_layer_31_cast_fp16")];
+            tensor<int32, [3]> var_3509 = const()[name = string("op_3509"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3505_cast_fp16 = transpose(perm = var_3504, x = x_311_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [16, 192, 64]> value_layer_31_cast_fp16 = reshape(shape = var_3509, x = var_3505_cast_fp16)[name = string("value_layer_31_cast_fp16")];
             tensor<int32, [3]> var_3519_perm_0 = const()[name = string("op_3519_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3521_y_0_to_fp16 = const()[name = string("_inversed_3521_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_3519_cast_fp16 = transpose(perm = var_3519_perm_0, x = key_layer_31_cast_fp16)[name = string("transpose_50")];
-            tensor<fp16, [16, 64, 256]> _inversed_3521_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_3521_y_0_to_fp16)[name = string("_inversed_3521_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_3519_cast_fp16 = transpose(perm = var_3519_perm_0, x = key_layer_31_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [16, 64, 192]> _inversed_3521_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_3521_y_0_to_fp16)[name = string("_inversed_3521_cast_fp16")];
             bool attention_scores_91_transpose_x_0 = const()[name = string("attention_scores_91_transpose_x_0"), val = bool(false)];
             bool attention_scores_91_transpose_y_0 = const()[name = string("attention_scores_91_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_91_cast_fp16 = matmul(transpose_x = attention_scores_91_transpose_x_0, transpose_y = attention_scores_91_transpose_y_0, x = query_layer_93_cast_fp16, y = _inversed_3521_cast_fp16)[name = string("attention_scores_91_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_91_cast_fp16 = matmul(transpose_x = attention_scores_91_transpose_x_0, transpose_y = attention_scores_91_transpose_y_0, x = query_layer_93_cast_fp16, y = _inversed_3521_cast_fp16)[name = string("attention_scores_91_cast_fp16")];
             bool c2p_att_61_transpose_x_0 = const()[name = string("c2p_att_61_transpose_x_0"), val = bool(false)];
             bool c2p_att_61_transpose_y_0 = const()[name = string("c2p_att_61_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3577_to_fp16 = const()[name = string("op_3577_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(678879424)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_61_cast_fp16 = matmul(transpose_x = c2p_att_61_transpose_x_0, transpose_y = c2p_att_61_transpose_y_0, x = query_layer_93_cast_fp16, y = var_3577_to_fp16)[name = string("c2p_att_61_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_61_cast_fp16 = matmul(transpose_x = c2p_att_61_transpose_x_0, transpose_y = c2p_att_61_transpose_y_0, x = query_layer_93_cast_fp16, y = var_3577_to_fp16)[name = string("c2p_att_61_cast_fp16")];
             bool c2p_att_63_validate_indices_0 = const()[name = string("c2p_att_63_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_63_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_63_validate_indices_0, x = c2p_att_61_cast_fp16)[name = string("c2p_att_63_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_63_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_63_validate_indices_0, x = c2p_att_61_cast_fp16)[name = string("c2p_att_63_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3590_y_0_to_fp16 = const()[name = string("_inversed_3590_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_3590_cast_fp16 = mul(x = c2p_att_63_cast_fp16_cast_int16, y = _inversed_3590_y_0_to_fp16)[name = string("_inversed_3590_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_3590_cast_fp16 = mul(x = c2p_att_63_cast_fp16_cast_int16, y = _inversed_3590_y_0_to_fp16)[name = string("_inversed_3590_cast_fp16")];
             bool p2c_att_61_transpose_x_0 = const()[name = string("p2c_att_61_transpose_x_0"), val = bool(false)];
             bool p2c_att_61_transpose_y_0 = const()[name = string("p2c_att_61_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3605_to_fp16 = const()[name = string("op_3605_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679928064)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_61_cast_fp16 = matmul(transpose_x = p2c_att_61_transpose_x_0, transpose_y = p2c_att_61_transpose_y_0, x = key_layer_31_cast_fp16, y = var_3605_to_fp16)[name = string("p2c_att_61_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_61_cast_fp16 = matmul(transpose_x = p2c_att_61_transpose_x_0, transpose_y = p2c_att_61_transpose_y_0, x = key_layer_31_cast_fp16, y = var_3605_to_fp16)[name = string("p2c_att_61_cast_fp16")];
             bool var_3613_validate_indices_0 = const()[name = string("op_3613_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_3613_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3613_validate_indices_0, x = p2c_att_61_cast_fp16)[name = string("op_3613_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_3613_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3613_validate_indices_0, x = p2c_att_61_cast_fp16)[name = string("op_3613_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_63_perm_0 = const()[name = string("p2c_att_63_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3616_y_0_to_fp16 = const()[name = string("_inversed_3616_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_63_cast_fp16 = transpose(perm = p2c_att_63_perm_0, x = var_3613_cast_fp16_cast_int16)[name = string("transpose_49")];
-            tensor<fp16, [16, 256, 256]> _inversed_3616_cast_fp16 = mul(x = p2c_att_63_cast_fp16, y = _inversed_3616_y_0_to_fp16)[name = string("_inversed_3616_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_31_cast_fp16 = add(x = _inversed_3590_cast_fp16, y = _inversed_3616_cast_fp16)[name = string("rel_att_31_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_93_cast_fp16 = add(x = attention_scores_91_cast_fp16, y = rel_att_31_cast_fp16)[name = string("attention_scores_93_cast_fp16")];
-            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_95_cast_fp16 = reshape(shape = var_3621, x = attention_scores_93_cast_fp16)[name = string("attention_scores_95_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_369_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_95_cast_fp16, cond = var_339)[name = string("input_369_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_371_cast_fp16 = softmax(axis = var_56, x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
-            tensor<int32, [3]> var_3630 = const()[name = string("op_3630"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_3631_cast_fp16 = reshape(shape = var_3630, x = input_371_cast_fp16)[name = string("op_3631_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_63_cast_fp16 = transpose(perm = p2c_att_63_perm_0, x = var_3613_cast_fp16_cast_int16)[name = string("transpose_49")];
+            tensor<fp16, [16, 192, 192]> _inversed_3616_cast_fp16 = mul(x = p2c_att_63_cast_fp16, y = _inversed_3616_y_0_to_fp16)[name = string("_inversed_3616_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_31_cast_fp16 = add(x = _inversed_3590_cast_fp16, y = _inversed_3616_cast_fp16)[name = string("rel_att_31_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_93_cast_fp16 = add(x = attention_scores_91_cast_fp16, y = rel_att_31_cast_fp16)[name = string("attention_scores_93_cast_fp16")];
+            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_95_cast_fp16 = reshape(shape = var_3621, x = attention_scores_93_cast_fp16)[name = string("attention_scores_95_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_369_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_95_cast_fp16, cond = var_339)[name = string("input_369_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_371_cast_fp16 = softmax(axis = var_56, x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            tensor<int32, [3]> var_3630 = const()[name = string("op_3630"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_3631_cast_fp16 = reshape(shape = var_3630, x = input_371_cast_fp16)[name = string("op_3631_cast_fp16")];
             bool context_layer_61_transpose_x_0 = const()[name = string("context_layer_61_transpose_x_0"), val = bool(false)];
             bool context_layer_61_transpose_y_0 = const()[name = string("context_layer_61_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_61_cast_fp16 = matmul(transpose_x = context_layer_61_transpose_x_0, transpose_y = context_layer_61_transpose_y_0, x = var_3631_cast_fp16, y = value_layer_31_cast_fp16)[name = string("context_layer_61_cast_fp16")];
-            tensor<int32, [4]> var_3635 = const()[name = string("op_3635"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3636_cast_fp16 = reshape(shape = var_3635, x = context_layer_61_cast_fp16)[name = string("op_3636_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_61_cast_fp16 = matmul(transpose_x = context_layer_61_transpose_x_0, transpose_y = context_layer_61_transpose_y_0, x = var_3631_cast_fp16, y = value_layer_31_cast_fp16)[name = string("context_layer_61_cast_fp16")];
+            tensor<int32, [4]> var_3635 = const()[name = string("op_3635"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3636_cast_fp16 = reshape(shape = var_3635, x = context_layer_61_cast_fp16)[name = string("op_3636_cast_fp16")];
             tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3642 = const()[name = string("op_3642"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_3638_cast_fp16 = transpose(perm = var_3637, x = var_3636_cast_fp16)[name = string("transpose_48")];
-            tensor<fp16, [1, 256, 1024]> input_373_cast_fp16 = reshape(shape = var_3642, x = var_3638_cast_fp16)[name = string("input_373_cast_fp16")];
+            tensor<int32, [3]> var_3642 = const()[name = string("op_3642"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_3638_cast_fp16 = transpose(perm = var_3637, x = var_3636_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 192, 1024]> input_373_cast_fp16 = reshape(shape = var_3642, x = var_3638_cast_fp16)[name = string("input_373_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680976704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683073920)))];
-            tensor<fp16, [1, 256, 1024]> linear_125_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_attention_output_dense_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_125_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_377_cast_fp16 = add(x = linear_125_cast_fp16, y = input_365_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_125_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_attention_output_dense_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_377_cast_fp16 = add(x = linear_125_cast_fp16, y = input_365_cast_fp16)[name = string("input_377_cast_fp16")];
             tensor<int32, [1]> input_379_axes_0 = const()[name = string("input_379_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683076032)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683078144)))];
-            tensor<fp16, [1, 256, 1024]> input_379_cast_fp16 = layer_norm(axes = input_379_axes_0, beta = encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16, x = input_377_cast_fp16)[name = string("input_379_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_379_cast_fp16 = layer_norm(axes = input_379_axes_0, beta = encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16, x = input_377_cast_fp16)[name = string("input_379_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_15_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683080256)))];
             tensor<fp16, [4096]> encoder_encoder_layer_15_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691468928)))];
-            tensor<fp16, [1, 256, 4096]> linear_126_cast_fp16 = linear(bias = encoder_encoder_layer_15_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_15_intermediate_dense_weight_to_fp16, x = input_379_cast_fp16)[name = string("linear_126_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_126_cast_fp16 = linear(bias = encoder_encoder_layer_15_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_15_intermediate_dense_weight_to_fp16, x = input_379_cast_fp16)[name = string("linear_126_cast_fp16")];
             string input_383_mode_0 = const()[name = string("input_383_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_383_cast_fp16 = gelu(mode = input_383_mode_0, x = linear_126_cast_fp16)[name = string("input_383_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_383_cast_fp16 = gelu(mode = input_383_mode_0, x = linear_126_cast_fp16)[name = string("input_383_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_15_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691477184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699865856)))];
-            tensor<fp16, [1, 256, 1024]> linear_127_cast_fp16 = linear(bias = encoder_encoder_layer_15_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_output_dense_weight_to_fp16, x = input_383_cast_fp16)[name = string("linear_127_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_387_cast_fp16 = add(x = linear_127_cast_fp16, y = input_379_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_127_cast_fp16 = linear(bias = encoder_encoder_layer_15_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_output_dense_weight_to_fp16, x = input_383_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_387_cast_fp16 = add(x = linear_127_cast_fp16, y = input_379_cast_fp16)[name = string("input_387_cast_fp16")];
             tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699867968)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699870080)))];
-            tensor<fp16, [1, 256, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699872192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(701969408)))];
-            tensor<fp16, [1, 256, 1024]> linear_128_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_128_cast_fp16")];
-            tensor<int32, [4]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_323_cast_fp16 = reshape(shape = var_3693, x = linear_128_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_128_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_128_cast_fp16")];
+            tensor<int32, [4]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_323_cast_fp16 = reshape(shape = var_3693, x = linear_128_cast_fp16)[name = string("x_323_cast_fp16")];
             tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3700 = const()[name = string("op_3700"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3696_cast_fp16 = transpose(perm = var_3695, x = x_323_cast_fp16)[name = string("transpose_47")];
-            tensor<fp16, [16, 256, 64]> query_layer_99_cast_fp16 = reshape(shape = var_3700, x = var_3696_cast_fp16)[name = string("query_layer_99_cast_fp16")];
+            tensor<int32, [3]> var_3700 = const()[name = string("op_3700"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3696_cast_fp16 = transpose(perm = var_3695, x = x_323_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [16, 192, 64]> query_layer_99_cast_fp16 = reshape(shape = var_3700, x = var_3696_cast_fp16)[name = string("query_layer_99_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(701971520)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704068736)))];
-            tensor<fp16, [1, 256, 1024]> linear_129_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_129_cast_fp16")];
-            tensor<int32, [4]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_327_cast_fp16 = reshape(shape = var_3707, x = linear_129_cast_fp16)[name = string("x_327_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_129_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_327_cast_fp16 = reshape(shape = var_3707, x = linear_129_cast_fp16)[name = string("x_327_cast_fp16")];
             tensor<int32, [4]> var_3709 = const()[name = string("op_3709"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3714 = const()[name = string("op_3714"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3710_cast_fp16 = transpose(perm = var_3709, x = x_327_cast_fp16)[name = string("transpose_46")];
-            tensor<fp16, [16, 256, 64]> key_layer_33_cast_fp16 = reshape(shape = var_3714, x = var_3710_cast_fp16)[name = string("key_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_3714 = const()[name = string("op_3714"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3710_cast_fp16 = transpose(perm = var_3709, x = x_327_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [16, 192, 64]> key_layer_33_cast_fp16 = reshape(shape = var_3714, x = var_3710_cast_fp16)[name = string("key_layer_33_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704070848)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706168064)))];
-            tensor<fp16, [1, 256, 1024]> linear_130_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_130_cast_fp16")];
-            tensor<int32, [4]> var_3721 = const()[name = string("op_3721"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_331_cast_fp16 = reshape(shape = var_3721, x = linear_130_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_130_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3721 = const()[name = string("op_3721"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_331_cast_fp16 = reshape(shape = var_3721, x = linear_130_cast_fp16)[name = string("x_331_cast_fp16")];
             tensor<int32, [4]> var_3723 = const()[name = string("op_3723"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3728 = const()[name = string("op_3728"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3724_cast_fp16 = transpose(perm = var_3723, x = x_331_cast_fp16)[name = string("transpose_45")];
-            tensor<fp16, [16, 256, 64]> value_layer_33_cast_fp16 = reshape(shape = var_3728, x = var_3724_cast_fp16)[name = string("value_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_3728 = const()[name = string("op_3728"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3724_cast_fp16 = transpose(perm = var_3723, x = x_331_cast_fp16)[name = string("transpose_45")];
+            tensor<fp16, [16, 192, 64]> value_layer_33_cast_fp16 = reshape(shape = var_3728, x = var_3724_cast_fp16)[name = string("value_layer_33_cast_fp16")];
             tensor<int32, [3]> var_3738_perm_0 = const()[name = string("op_3738_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3740_y_0_to_fp16 = const()[name = string("_inversed_3740_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_3738_cast_fp16 = transpose(perm = var_3738_perm_0, x = key_layer_33_cast_fp16)[name = string("transpose_44")];
-            tensor<fp16, [16, 64, 256]> _inversed_3740_cast_fp16 = mul(x = var_3738_cast_fp16, y = _inversed_3740_y_0_to_fp16)[name = string("_inversed_3740_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_3738_cast_fp16 = transpose(perm = var_3738_perm_0, x = key_layer_33_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [16, 64, 192]> _inversed_3740_cast_fp16 = mul(x = var_3738_cast_fp16, y = _inversed_3740_y_0_to_fp16)[name = string("_inversed_3740_cast_fp16")];
             bool attention_scores_97_transpose_x_0 = const()[name = string("attention_scores_97_transpose_x_0"), val = bool(false)];
             bool attention_scores_97_transpose_y_0 = const()[name = string("attention_scores_97_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_97_cast_fp16 = matmul(transpose_x = attention_scores_97_transpose_x_0, transpose_y = attention_scores_97_transpose_y_0, x = query_layer_99_cast_fp16, y = _inversed_3740_cast_fp16)[name = string("attention_scores_97_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_97_cast_fp16 = matmul(transpose_x = attention_scores_97_transpose_x_0, transpose_y = attention_scores_97_transpose_y_0, x = query_layer_99_cast_fp16, y = _inversed_3740_cast_fp16)[name = string("attention_scores_97_cast_fp16")];
             bool c2p_att_65_transpose_x_0 = const()[name = string("c2p_att_65_transpose_x_0"), val = bool(false)];
             bool c2p_att_65_transpose_y_0 = const()[name = string("c2p_att_65_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3796_to_fp16 = const()[name = string("op_3796_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706170176)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_65_cast_fp16 = matmul(transpose_x = c2p_att_65_transpose_x_0, transpose_y = c2p_att_65_transpose_y_0, x = query_layer_99_cast_fp16, y = var_3796_to_fp16)[name = string("c2p_att_65_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_65_cast_fp16 = matmul(transpose_x = c2p_att_65_transpose_x_0, transpose_y = c2p_att_65_transpose_y_0, x = query_layer_99_cast_fp16, y = var_3796_to_fp16)[name = string("c2p_att_65_cast_fp16")];
             bool c2p_att_67_validate_indices_0 = const()[name = string("c2p_att_67_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_67_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_67_validate_indices_0, x = c2p_att_65_cast_fp16)[name = string("c2p_att_67_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_67_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_67_validate_indices_0, x = c2p_att_65_cast_fp16)[name = string("c2p_att_67_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3809_y_0_to_fp16 = const()[name = string("_inversed_3809_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_3809_cast_fp16 = mul(x = c2p_att_67_cast_fp16_cast_int16, y = _inversed_3809_y_0_to_fp16)[name = string("_inversed_3809_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_3809_cast_fp16 = mul(x = c2p_att_67_cast_fp16_cast_int16, y = _inversed_3809_y_0_to_fp16)[name = string("_inversed_3809_cast_fp16")];
             bool p2c_att_65_transpose_x_0 = const()[name = string("p2c_att_65_transpose_x_0"), val = bool(false)];
             bool p2c_att_65_transpose_y_0 = const()[name = string("p2c_att_65_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3824_to_fp16 = const()[name = string("op_3824_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707218816)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_65_cast_fp16 = matmul(transpose_x = p2c_att_65_transpose_x_0, transpose_y = p2c_att_65_transpose_y_0, x = key_layer_33_cast_fp16, y = var_3824_to_fp16)[name = string("p2c_att_65_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_65_cast_fp16 = matmul(transpose_x = p2c_att_65_transpose_x_0, transpose_y = p2c_att_65_transpose_y_0, x = key_layer_33_cast_fp16, y = var_3824_to_fp16)[name = string("p2c_att_65_cast_fp16")];
             bool var_3832_validate_indices_0 = const()[name = string("op_3832_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_3832_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3832_validate_indices_0, x = p2c_att_65_cast_fp16)[name = string("op_3832_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_3832_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3832_validate_indices_0, x = p2c_att_65_cast_fp16)[name = string("op_3832_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_67_perm_0 = const()[name = string("p2c_att_67_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3835_y_0_to_fp16 = const()[name = string("_inversed_3835_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_67_cast_fp16 = transpose(perm = p2c_att_67_perm_0, x = var_3832_cast_fp16_cast_int16)[name = string("transpose_43")];
-            tensor<fp16, [16, 256, 256]> _inversed_3835_cast_fp16 = mul(x = p2c_att_67_cast_fp16, y = _inversed_3835_y_0_to_fp16)[name = string("_inversed_3835_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_33_cast_fp16 = add(x = _inversed_3809_cast_fp16, y = _inversed_3835_cast_fp16)[name = string("rel_att_33_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_99_cast_fp16 = add(x = attention_scores_97_cast_fp16, y = rel_att_33_cast_fp16)[name = string("attention_scores_99_cast_fp16")];
-            tensor<int32, [4]> var_3840 = const()[name = string("op_3840"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_101_cast_fp16 = reshape(shape = var_3840, x = attention_scores_99_cast_fp16)[name = string("attention_scores_101_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_393_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_101_cast_fp16, cond = var_339)[name = string("input_393_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_395_cast_fp16 = softmax(axis = var_56, x = input_393_cast_fp16)[name = string("input_395_cast_fp16")];
-            tensor<int32, [3]> var_3849 = const()[name = string("op_3849"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_3850_cast_fp16 = reshape(shape = var_3849, x = input_395_cast_fp16)[name = string("op_3850_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_67_cast_fp16 = transpose(perm = p2c_att_67_perm_0, x = var_3832_cast_fp16_cast_int16)[name = string("transpose_43")];
+            tensor<fp16, [16, 192, 192]> _inversed_3835_cast_fp16 = mul(x = p2c_att_67_cast_fp16, y = _inversed_3835_y_0_to_fp16)[name = string("_inversed_3835_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_33_cast_fp16 = add(x = _inversed_3809_cast_fp16, y = _inversed_3835_cast_fp16)[name = string("rel_att_33_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_99_cast_fp16 = add(x = attention_scores_97_cast_fp16, y = rel_att_33_cast_fp16)[name = string("attention_scores_99_cast_fp16")];
+            tensor<int32, [4]> var_3840 = const()[name = string("op_3840"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_101_cast_fp16 = reshape(shape = var_3840, x = attention_scores_99_cast_fp16)[name = string("attention_scores_101_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_393_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_101_cast_fp16, cond = var_339)[name = string("input_393_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_395_cast_fp16 = softmax(axis = var_56, x = input_393_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<int32, [3]> var_3849 = const()[name = string("op_3849"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_3850_cast_fp16 = reshape(shape = var_3849, x = input_395_cast_fp16)[name = string("op_3850_cast_fp16")];
             bool context_layer_65_transpose_x_0 = const()[name = string("context_layer_65_transpose_x_0"), val = bool(false)];
             bool context_layer_65_transpose_y_0 = const()[name = string("context_layer_65_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_65_cast_fp16 = matmul(transpose_x = context_layer_65_transpose_x_0, transpose_y = context_layer_65_transpose_y_0, x = var_3850_cast_fp16, y = value_layer_33_cast_fp16)[name = string("context_layer_65_cast_fp16")];
-            tensor<int32, [4]> var_3854 = const()[name = string("op_3854"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3855_cast_fp16 = reshape(shape = var_3854, x = context_layer_65_cast_fp16)[name = string("op_3855_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_65_cast_fp16 = matmul(transpose_x = context_layer_65_transpose_x_0, transpose_y = context_layer_65_transpose_y_0, x = var_3850_cast_fp16, y = value_layer_33_cast_fp16)[name = string("context_layer_65_cast_fp16")];
+            tensor<int32, [4]> var_3854 = const()[name = string("op_3854"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3855_cast_fp16 = reshape(shape = var_3854, x = context_layer_65_cast_fp16)[name = string("op_3855_cast_fp16")];
             tensor<int32, [4]> var_3856 = const()[name = string("op_3856"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_3857_cast_fp16 = transpose(perm = var_3856, x = var_3855_cast_fp16)[name = string("transpose_42")];
-            tensor<fp16, [1, 256, 1024]> input_397_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_397_cast_fp16")];
+            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_3857_cast_fp16 = transpose(perm = var_3856, x = var_3855_cast_fp16)[name = string("transpose_42")];
+            tensor<fp16, [1, 192, 1024]> input_397_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_397_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708267456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710364672)))];
-            tensor<fp16, [1, 256, 1024]> linear_133_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_attention_output_dense_weight_to_fp16, x = input_397_cast_fp16)[name = string("linear_133_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_401_cast_fp16 = add(x = linear_133_cast_fp16, y = input_389_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_133_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_attention_output_dense_weight_to_fp16, x = input_397_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_401_cast_fp16 = add(x = linear_133_cast_fp16, y = input_389_cast_fp16)[name = string("input_401_cast_fp16")];
             tensor<int32, [1]> input_403_axes_0 = const()[name = string("input_403_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710366784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710368896)))];
-            tensor<fp16, [1, 256, 1024]> input_403_cast_fp16 = layer_norm(axes = input_403_axes_0, beta = encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16, x = input_401_cast_fp16)[name = string("input_403_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_403_cast_fp16 = layer_norm(axes = input_403_axes_0, beta = encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16, x = input_401_cast_fp16)[name = string("input_403_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_16_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710371008)))];
             tensor<fp16, [4096]> encoder_encoder_layer_16_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(718759680)))];
-            tensor<fp16, [1, 256, 4096]> linear_134_cast_fp16 = linear(bias = encoder_encoder_layer_16_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_16_intermediate_dense_weight_to_fp16, x = input_403_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_134_cast_fp16 = linear(bias = encoder_encoder_layer_16_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_16_intermediate_dense_weight_to_fp16, x = input_403_cast_fp16)[name = string("linear_134_cast_fp16")];
             string input_407_mode_0 = const()[name = string("input_407_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_407_cast_fp16 = gelu(mode = input_407_mode_0, x = linear_134_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_407_cast_fp16 = gelu(mode = input_407_mode_0, x = linear_134_cast_fp16)[name = string("input_407_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_16_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(718767936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727156608)))];
-            tensor<fp16, [1, 256, 1024]> linear_135_cast_fp16 = linear(bias = encoder_encoder_layer_16_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_output_dense_weight_to_fp16, x = input_407_cast_fp16)[name = string("linear_135_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_411_cast_fp16 = add(x = linear_135_cast_fp16, y = input_403_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_135_cast_fp16 = linear(bias = encoder_encoder_layer_16_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_output_dense_weight_to_fp16, x = input_407_cast_fp16)[name = string("linear_135_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_411_cast_fp16 = add(x = linear_135_cast_fp16, y = input_403_cast_fp16)[name = string("input_411_cast_fp16")];
             tensor<int32, [1]> input_413_axes_0 = const()[name = string("input_413_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727158720)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727160832)))];
-            tensor<fp16, [1, 256, 1024]> input_413_cast_fp16 = layer_norm(axes = input_413_axes_0, beta = encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16, x = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_413_cast_fp16 = layer_norm(axes = input_413_axes_0, beta = encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16, x = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727162944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729260160)))];
-            tensor<fp16, [1, 256, 1024]> linear_136_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_136_cast_fp16")];
-            tensor<int32, [4]> var_3912 = const()[name = string("op_3912"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_343_cast_fp16 = reshape(shape = var_3912, x = linear_136_cast_fp16)[name = string("x_343_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_136_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<int32, [4]> var_3912 = const()[name = string("op_3912"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_343_cast_fp16 = reshape(shape = var_3912, x = linear_136_cast_fp16)[name = string("x_343_cast_fp16")];
             tensor<int32, [4]> var_3914 = const()[name = string("op_3914"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3919 = const()[name = string("op_3919"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3915_cast_fp16 = transpose(perm = var_3914, x = x_343_cast_fp16)[name = string("transpose_41")];
-            tensor<fp16, [16, 256, 64]> query_layer_105_cast_fp16 = reshape(shape = var_3919, x = var_3915_cast_fp16)[name = string("query_layer_105_cast_fp16")];
+            tensor<int32, [3]> var_3919 = const()[name = string("op_3919"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3915_cast_fp16 = transpose(perm = var_3914, x = x_343_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [16, 192, 64]> query_layer_105_cast_fp16 = reshape(shape = var_3919, x = var_3915_cast_fp16)[name = string("query_layer_105_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729262272)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731359488)))];
-            tensor<fp16, [1, 256, 1024]> linear_137_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_137_cast_fp16")];
-            tensor<int32, [4]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_347_cast_fp16 = reshape(shape = var_3926, x = linear_137_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_137_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_137_cast_fp16")];
+            tensor<int32, [4]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_347_cast_fp16 = reshape(shape = var_3926, x = linear_137_cast_fp16)[name = string("x_347_cast_fp16")];
             tensor<int32, [4]> var_3928 = const()[name = string("op_3928"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3929_cast_fp16 = transpose(perm = var_3928, x = x_347_cast_fp16)[name = string("transpose_40")];
-            tensor<fp16, [16, 256, 64]> key_layer_35_cast_fp16 = reshape(shape = var_3933, x = var_3929_cast_fp16)[name = string("key_layer_35_cast_fp16")];
+            tensor<int32, [3]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3929_cast_fp16 = transpose(perm = var_3928, x = x_347_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [16, 192, 64]> key_layer_35_cast_fp16 = reshape(shape = var_3933, x = var_3929_cast_fp16)[name = string("key_layer_35_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731361600)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733458816)))];
-            tensor<fp16, [1, 256, 1024]> linear_138_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_138_cast_fp16")];
-            tensor<int32, [4]> var_3940 = const()[name = string("op_3940"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_351_cast_fp16 = reshape(shape = var_3940, x = linear_138_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_138_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3940 = const()[name = string("op_3940"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_351_cast_fp16 = reshape(shape = var_3940, x = linear_138_cast_fp16)[name = string("x_351_cast_fp16")];
             tensor<int32, [4]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3947 = const()[name = string("op_3947"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_3943_cast_fp16 = transpose(perm = var_3942, x = x_351_cast_fp16)[name = string("transpose_39")];
-            tensor<fp16, [16, 256, 64]> value_layer_35_cast_fp16 = reshape(shape = var_3947, x = var_3943_cast_fp16)[name = string("value_layer_35_cast_fp16")];
+            tensor<int32, [3]> var_3947 = const()[name = string("op_3947"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_3943_cast_fp16 = transpose(perm = var_3942, x = x_351_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [16, 192, 64]> value_layer_35_cast_fp16 = reshape(shape = var_3947, x = var_3943_cast_fp16)[name = string("value_layer_35_cast_fp16")];
             tensor<int32, [3]> var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3959_y_0_to_fp16 = const()[name = string("_inversed_3959_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = key_layer_35_cast_fp16)[name = string("transpose_38")];
-            tensor<fp16, [16, 64, 256]> _inversed_3959_cast_fp16 = mul(x = var_3957_cast_fp16, y = _inversed_3959_y_0_to_fp16)[name = string("_inversed_3959_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = key_layer_35_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [16, 64, 192]> _inversed_3959_cast_fp16 = mul(x = var_3957_cast_fp16, y = _inversed_3959_y_0_to_fp16)[name = string("_inversed_3959_cast_fp16")];
             bool attention_scores_103_transpose_x_0 = const()[name = string("attention_scores_103_transpose_x_0"), val = bool(false)];
             bool attention_scores_103_transpose_y_0 = const()[name = string("attention_scores_103_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_103_cast_fp16 = matmul(transpose_x = attention_scores_103_transpose_x_0, transpose_y = attention_scores_103_transpose_y_0, x = query_layer_105_cast_fp16, y = _inversed_3959_cast_fp16)[name = string("attention_scores_103_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_103_cast_fp16 = matmul(transpose_x = attention_scores_103_transpose_x_0, transpose_y = attention_scores_103_transpose_y_0, x = query_layer_105_cast_fp16, y = _inversed_3959_cast_fp16)[name = string("attention_scores_103_cast_fp16")];
             bool c2p_att_69_transpose_x_0 = const()[name = string("c2p_att_69_transpose_x_0"), val = bool(false)];
             bool c2p_att_69_transpose_y_0 = const()[name = string("c2p_att_69_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4015_to_fp16 = const()[name = string("op_4015_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733460928)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_69_cast_fp16 = matmul(transpose_x = c2p_att_69_transpose_x_0, transpose_y = c2p_att_69_transpose_y_0, x = query_layer_105_cast_fp16, y = var_4015_to_fp16)[name = string("c2p_att_69_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_69_cast_fp16 = matmul(transpose_x = c2p_att_69_transpose_x_0, transpose_y = c2p_att_69_transpose_y_0, x = query_layer_105_cast_fp16, y = var_4015_to_fp16)[name = string("c2p_att_69_cast_fp16")];
             bool c2p_att_71_validate_indices_0 = const()[name = string("c2p_att_71_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_71_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_71_validate_indices_0, x = c2p_att_69_cast_fp16)[name = string("c2p_att_71_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_71_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_71_validate_indices_0, x = c2p_att_69_cast_fp16)[name = string("c2p_att_71_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4028_y_0_to_fp16 = const()[name = string("_inversed_4028_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_4028_cast_fp16 = mul(x = c2p_att_71_cast_fp16_cast_int16, y = _inversed_4028_y_0_to_fp16)[name = string("_inversed_4028_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_4028_cast_fp16 = mul(x = c2p_att_71_cast_fp16_cast_int16, y = _inversed_4028_y_0_to_fp16)[name = string("_inversed_4028_cast_fp16")];
             bool p2c_att_69_transpose_x_0 = const()[name = string("p2c_att_69_transpose_x_0"), val = bool(false)];
             bool p2c_att_69_transpose_y_0 = const()[name = string("p2c_att_69_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734509568)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_69_cast_fp16 = matmul(transpose_x = p2c_att_69_transpose_x_0, transpose_y = p2c_att_69_transpose_y_0, x = key_layer_35_cast_fp16, y = var_4043_to_fp16)[name = string("p2c_att_69_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_69_cast_fp16 = matmul(transpose_x = p2c_att_69_transpose_x_0, transpose_y = p2c_att_69_transpose_y_0, x = key_layer_35_cast_fp16, y = var_4043_to_fp16)[name = string("p2c_att_69_cast_fp16")];
             bool var_4051_validate_indices_0 = const()[name = string("op_4051_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_4051_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4051_validate_indices_0, x = p2c_att_69_cast_fp16)[name = string("op_4051_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_4051_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4051_validate_indices_0, x = p2c_att_69_cast_fp16)[name = string("op_4051_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_71_perm_0 = const()[name = string("p2c_att_71_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4054_y_0_to_fp16 = const()[name = string("_inversed_4054_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_71_cast_fp16 = transpose(perm = p2c_att_71_perm_0, x = var_4051_cast_fp16_cast_int16)[name = string("transpose_37")];
-            tensor<fp16, [16, 256, 256]> _inversed_4054_cast_fp16 = mul(x = p2c_att_71_cast_fp16, y = _inversed_4054_y_0_to_fp16)[name = string("_inversed_4054_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_35_cast_fp16 = add(x = _inversed_4028_cast_fp16, y = _inversed_4054_cast_fp16)[name = string("rel_att_35_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_105_cast_fp16 = add(x = attention_scores_103_cast_fp16, y = rel_att_35_cast_fp16)[name = string("attention_scores_105_cast_fp16")];
-            tensor<int32, [4]> var_4059 = const()[name = string("op_4059"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_107_cast_fp16 = reshape(shape = var_4059, x = attention_scores_105_cast_fp16)[name = string("attention_scores_107_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_417_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_107_cast_fp16, cond = var_339)[name = string("input_417_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_419_cast_fp16 = softmax(axis = var_56, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
-            tensor<int32, [3]> var_4068 = const()[name = string("op_4068"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_4069_cast_fp16 = reshape(shape = var_4068, x = input_419_cast_fp16)[name = string("op_4069_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_71_cast_fp16 = transpose(perm = p2c_att_71_perm_0, x = var_4051_cast_fp16_cast_int16)[name = string("transpose_37")];
+            tensor<fp16, [16, 192, 192]> _inversed_4054_cast_fp16 = mul(x = p2c_att_71_cast_fp16, y = _inversed_4054_y_0_to_fp16)[name = string("_inversed_4054_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_35_cast_fp16 = add(x = _inversed_4028_cast_fp16, y = _inversed_4054_cast_fp16)[name = string("rel_att_35_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_105_cast_fp16 = add(x = attention_scores_103_cast_fp16, y = rel_att_35_cast_fp16)[name = string("attention_scores_105_cast_fp16")];
+            tensor<int32, [4]> var_4059 = const()[name = string("op_4059"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_107_cast_fp16 = reshape(shape = var_4059, x = attention_scores_105_cast_fp16)[name = string("attention_scores_107_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_417_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_107_cast_fp16, cond = var_339)[name = string("input_417_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_419_cast_fp16 = softmax(axis = var_56, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
+            tensor<int32, [3]> var_4068 = const()[name = string("op_4068"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_4069_cast_fp16 = reshape(shape = var_4068, x = input_419_cast_fp16)[name = string("op_4069_cast_fp16")];
             bool context_layer_69_transpose_x_0 = const()[name = string("context_layer_69_transpose_x_0"), val = bool(false)];
             bool context_layer_69_transpose_y_0 = const()[name = string("context_layer_69_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_69_cast_fp16 = matmul(transpose_x = context_layer_69_transpose_x_0, transpose_y = context_layer_69_transpose_y_0, x = var_4069_cast_fp16, y = value_layer_35_cast_fp16)[name = string("context_layer_69_cast_fp16")];
-            tensor<int32, [4]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4074_cast_fp16 = reshape(shape = var_4073, x = context_layer_69_cast_fp16)[name = string("op_4074_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_69_cast_fp16 = matmul(transpose_x = context_layer_69_transpose_x_0, transpose_y = context_layer_69_transpose_y_0, x = var_4069_cast_fp16, y = value_layer_35_cast_fp16)[name = string("context_layer_69_cast_fp16")];
+            tensor<int32, [4]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4074_cast_fp16 = reshape(shape = var_4073, x = context_layer_69_cast_fp16)[name = string("op_4074_cast_fp16")];
             tensor<int32, [4]> var_4075 = const()[name = string("op_4075"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_4076_cast_fp16 = transpose(perm = var_4075, x = var_4074_cast_fp16)[name = string("transpose_36")];
-            tensor<fp16, [1, 256, 1024]> input_421_cast_fp16 = reshape(shape = var_4080, x = var_4076_cast_fp16)[name = string("input_421_cast_fp16")];
+            tensor<int32, [3]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_4076_cast_fp16 = transpose(perm = var_4075, x = var_4074_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 192, 1024]> input_421_cast_fp16 = reshape(shape = var_4080, x = var_4076_cast_fp16)[name = string("input_421_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(735558208)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737655424)))];
-            tensor<fp16, [1, 256, 1024]> linear_141_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_attention_output_dense_weight_to_fp16, x = input_421_cast_fp16)[name = string("linear_141_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_425_cast_fp16 = add(x = linear_141_cast_fp16, y = input_413_cast_fp16)[name = string("input_425_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_141_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_attention_output_dense_weight_to_fp16, x = input_421_cast_fp16)[name = string("linear_141_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_425_cast_fp16 = add(x = linear_141_cast_fp16, y = input_413_cast_fp16)[name = string("input_425_cast_fp16")];
             tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737657536)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737659648)))];
-            tensor<fp16, [1, 256, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, beta = encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, beta = encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_17_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737661760)))];
             tensor<fp16, [4096]> encoder_encoder_layer_17_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746050432)))];
-            tensor<fp16, [1, 256, 4096]> linear_142_cast_fp16 = linear(bias = encoder_encoder_layer_17_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_17_intermediate_dense_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_142_cast_fp16 = linear(bias = encoder_encoder_layer_17_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_17_intermediate_dense_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_142_cast_fp16")];
             string input_431_mode_0 = const()[name = string("input_431_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_431_cast_fp16 = gelu(mode = input_431_mode_0, x = linear_142_cast_fp16)[name = string("input_431_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_431_cast_fp16 = gelu(mode = input_431_mode_0, x = linear_142_cast_fp16)[name = string("input_431_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_17_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746058688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754447360)))];
-            tensor<fp16, [1, 256, 1024]> linear_143_cast_fp16 = linear(bias = encoder_encoder_layer_17_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_output_dense_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_143_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_435_cast_fp16 = add(x = linear_143_cast_fp16, y = input_427_cast_fp16)[name = string("input_435_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_143_cast_fp16 = linear(bias = encoder_encoder_layer_17_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_output_dense_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_435_cast_fp16 = add(x = linear_143_cast_fp16, y = input_427_cast_fp16)[name = string("input_435_cast_fp16")];
             tensor<int32, [1]> input_437_axes_0 = const()[name = string("input_437_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754449472)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754451584)))];
-            tensor<fp16, [1, 256, 1024]> input_437_cast_fp16 = layer_norm(axes = input_437_axes_0, beta = encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16, x = input_435_cast_fp16)[name = string("input_437_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_437_cast_fp16 = layer_norm(axes = input_437_axes_0, beta = encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16, x = input_435_cast_fp16)[name = string("input_437_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754453696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(756550912)))];
-            tensor<fp16, [1, 256, 1024]> linear_144_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_144_cast_fp16")];
-            tensor<int32, [4]> var_4131 = const()[name = string("op_4131"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_363_cast_fp16 = reshape(shape = var_4131, x = linear_144_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_144_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_144_cast_fp16")];
+            tensor<int32, [4]> var_4131 = const()[name = string("op_4131"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_363_cast_fp16 = reshape(shape = var_4131, x = linear_144_cast_fp16)[name = string("x_363_cast_fp16")];
             tensor<int32, [4]> var_4133 = const()[name = string("op_4133"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4138 = const()[name = string("op_4138"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4134_cast_fp16 = transpose(perm = var_4133, x = x_363_cast_fp16)[name = string("transpose_35")];
-            tensor<fp16, [16, 256, 64]> query_layer_111_cast_fp16 = reshape(shape = var_4138, x = var_4134_cast_fp16)[name = string("query_layer_111_cast_fp16")];
+            tensor<int32, [3]> var_4138 = const()[name = string("op_4138"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4134_cast_fp16 = transpose(perm = var_4133, x = x_363_cast_fp16)[name = string("transpose_35")];
+            tensor<fp16, [16, 192, 64]> query_layer_111_cast_fp16 = reshape(shape = var_4138, x = var_4134_cast_fp16)[name = string("query_layer_111_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(756553024)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758650240)))];
-            tensor<fp16, [1, 256, 1024]> linear_145_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_145_cast_fp16")];
-            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_367_cast_fp16 = reshape(shape = var_4145, x = linear_145_cast_fp16)[name = string("x_367_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_145_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_367_cast_fp16 = reshape(shape = var_4145, x = linear_145_cast_fp16)[name = string("x_367_cast_fp16")];
             tensor<int32, [4]> var_4147 = const()[name = string("op_4147"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4152 = const()[name = string("op_4152"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4148_cast_fp16 = transpose(perm = var_4147, x = x_367_cast_fp16)[name = string("transpose_34")];
-            tensor<fp16, [16, 256, 64]> key_layer_37_cast_fp16 = reshape(shape = var_4152, x = var_4148_cast_fp16)[name = string("key_layer_37_cast_fp16")];
+            tensor<int32, [3]> var_4152 = const()[name = string("op_4152"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4148_cast_fp16 = transpose(perm = var_4147, x = x_367_cast_fp16)[name = string("transpose_34")];
+            tensor<fp16, [16, 192, 64]> key_layer_37_cast_fp16 = reshape(shape = var_4152, x = var_4148_cast_fp16)[name = string("key_layer_37_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758652352)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(760749568)))];
-            tensor<fp16, [1, 256, 1024]> linear_146_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_146_cast_fp16")];
-            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_371_cast_fp16 = reshape(shape = var_4159, x = linear_146_cast_fp16)[name = string("x_371_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_146_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_146_cast_fp16")];
+            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_371_cast_fp16 = reshape(shape = var_4159, x = linear_146_cast_fp16)[name = string("x_371_cast_fp16")];
             tensor<int32, [4]> var_4161 = const()[name = string("op_4161"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4166 = const()[name = string("op_4166"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4162_cast_fp16 = transpose(perm = var_4161, x = x_371_cast_fp16)[name = string("transpose_33")];
-            tensor<fp16, [16, 256, 64]> value_layer_37_cast_fp16 = reshape(shape = var_4166, x = var_4162_cast_fp16)[name = string("value_layer_37_cast_fp16")];
+            tensor<int32, [3]> var_4166 = const()[name = string("op_4166"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4162_cast_fp16 = transpose(perm = var_4161, x = x_371_cast_fp16)[name = string("transpose_33")];
+            tensor<fp16, [16, 192, 64]> value_layer_37_cast_fp16 = reshape(shape = var_4166, x = var_4162_cast_fp16)[name = string("value_layer_37_cast_fp16")];
             tensor<int32, [3]> var_4176_perm_0 = const()[name = string("op_4176_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4178_y_0_to_fp16 = const()[name = string("_inversed_4178_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_4176_cast_fp16 = transpose(perm = var_4176_perm_0, x = key_layer_37_cast_fp16)[name = string("transpose_32")];
-            tensor<fp16, [16, 64, 256]> _inversed_4178_cast_fp16 = mul(x = var_4176_cast_fp16, y = _inversed_4178_y_0_to_fp16)[name = string("_inversed_4178_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_4176_cast_fp16 = transpose(perm = var_4176_perm_0, x = key_layer_37_cast_fp16)[name = string("transpose_32")];
+            tensor<fp16, [16, 64, 192]> _inversed_4178_cast_fp16 = mul(x = var_4176_cast_fp16, y = _inversed_4178_y_0_to_fp16)[name = string("_inversed_4178_cast_fp16")];
             bool attention_scores_109_transpose_x_0 = const()[name = string("attention_scores_109_transpose_x_0"), val = bool(false)];
             bool attention_scores_109_transpose_y_0 = const()[name = string("attention_scores_109_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_109_cast_fp16 = matmul(transpose_x = attention_scores_109_transpose_x_0, transpose_y = attention_scores_109_transpose_y_0, x = query_layer_111_cast_fp16, y = _inversed_4178_cast_fp16)[name = string("attention_scores_109_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_109_cast_fp16 = matmul(transpose_x = attention_scores_109_transpose_x_0, transpose_y = attention_scores_109_transpose_y_0, x = query_layer_111_cast_fp16, y = _inversed_4178_cast_fp16)[name = string("attention_scores_109_cast_fp16")];
             bool c2p_att_73_transpose_x_0 = const()[name = string("c2p_att_73_transpose_x_0"), val = bool(false)];
             bool c2p_att_73_transpose_y_0 = const()[name = string("c2p_att_73_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(760751680)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_73_cast_fp16 = matmul(transpose_x = c2p_att_73_transpose_x_0, transpose_y = c2p_att_73_transpose_y_0, x = query_layer_111_cast_fp16, y = var_4234_to_fp16)[name = string("c2p_att_73_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_73_cast_fp16 = matmul(transpose_x = c2p_att_73_transpose_x_0, transpose_y = c2p_att_73_transpose_y_0, x = query_layer_111_cast_fp16, y = var_4234_to_fp16)[name = string("c2p_att_73_cast_fp16")];
             bool c2p_att_75_validate_indices_0 = const()[name = string("c2p_att_75_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_75_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_75_validate_indices_0, x = c2p_att_73_cast_fp16)[name = string("c2p_att_75_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_75_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_75_validate_indices_0, x = c2p_att_73_cast_fp16)[name = string("c2p_att_75_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4247_y_0_to_fp16 = const()[name = string("_inversed_4247_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_4247_cast_fp16 = mul(x = c2p_att_75_cast_fp16_cast_int16, y = _inversed_4247_y_0_to_fp16)[name = string("_inversed_4247_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_4247_cast_fp16 = mul(x = c2p_att_75_cast_fp16_cast_int16, y = _inversed_4247_y_0_to_fp16)[name = string("_inversed_4247_cast_fp16")];
             bool p2c_att_73_transpose_x_0 = const()[name = string("p2c_att_73_transpose_x_0"), val = bool(false)];
             bool p2c_att_73_transpose_y_0 = const()[name = string("p2c_att_73_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4262_to_fp16 = const()[name = string("op_4262_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761800320)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_73_cast_fp16 = matmul(transpose_x = p2c_att_73_transpose_x_0, transpose_y = p2c_att_73_transpose_y_0, x = key_layer_37_cast_fp16, y = var_4262_to_fp16)[name = string("p2c_att_73_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_73_cast_fp16 = matmul(transpose_x = p2c_att_73_transpose_x_0, transpose_y = p2c_att_73_transpose_y_0, x = key_layer_37_cast_fp16, y = var_4262_to_fp16)[name = string("p2c_att_73_cast_fp16")];
             bool var_4270_validate_indices_0 = const()[name = string("op_4270_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_4270_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4270_validate_indices_0, x = p2c_att_73_cast_fp16)[name = string("op_4270_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_4270_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4270_validate_indices_0, x = p2c_att_73_cast_fp16)[name = string("op_4270_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_75_perm_0 = const()[name = string("p2c_att_75_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4273_y_0_to_fp16 = const()[name = string("_inversed_4273_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_75_cast_fp16 = transpose(perm = p2c_att_75_perm_0, x = var_4270_cast_fp16_cast_int16)[name = string("transpose_31")];
-            tensor<fp16, [16, 256, 256]> _inversed_4273_cast_fp16 = mul(x = p2c_att_75_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = string("_inversed_4273_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_37_cast_fp16 = add(x = _inversed_4247_cast_fp16, y = _inversed_4273_cast_fp16)[name = string("rel_att_37_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_111_cast_fp16 = add(x = attention_scores_109_cast_fp16, y = rel_att_37_cast_fp16)[name = string("attention_scores_111_cast_fp16")];
-            tensor<int32, [4]> var_4278 = const()[name = string("op_4278"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_113_cast_fp16 = reshape(shape = var_4278, x = attention_scores_111_cast_fp16)[name = string("attention_scores_113_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_441_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_113_cast_fp16, cond = var_339)[name = string("input_441_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_443_cast_fp16 = softmax(axis = var_56, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
-            tensor<int32, [3]> var_4287 = const()[name = string("op_4287"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_4288_cast_fp16 = reshape(shape = var_4287, x = input_443_cast_fp16)[name = string("op_4288_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_75_cast_fp16 = transpose(perm = p2c_att_75_perm_0, x = var_4270_cast_fp16_cast_int16)[name = string("transpose_31")];
+            tensor<fp16, [16, 192, 192]> _inversed_4273_cast_fp16 = mul(x = p2c_att_75_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = string("_inversed_4273_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_37_cast_fp16 = add(x = _inversed_4247_cast_fp16, y = _inversed_4273_cast_fp16)[name = string("rel_att_37_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_111_cast_fp16 = add(x = attention_scores_109_cast_fp16, y = rel_att_37_cast_fp16)[name = string("attention_scores_111_cast_fp16")];
+            tensor<int32, [4]> var_4278 = const()[name = string("op_4278"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_113_cast_fp16 = reshape(shape = var_4278, x = attention_scores_111_cast_fp16)[name = string("attention_scores_113_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_441_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_113_cast_fp16, cond = var_339)[name = string("input_441_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_443_cast_fp16 = softmax(axis = var_56, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<int32, [3]> var_4287 = const()[name = string("op_4287"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_4288_cast_fp16 = reshape(shape = var_4287, x = input_443_cast_fp16)[name = string("op_4288_cast_fp16")];
             bool context_layer_73_transpose_x_0 = const()[name = string("context_layer_73_transpose_x_0"), val = bool(false)];
             bool context_layer_73_transpose_y_0 = const()[name = string("context_layer_73_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_73_cast_fp16 = matmul(transpose_x = context_layer_73_transpose_x_0, transpose_y = context_layer_73_transpose_y_0, x = var_4288_cast_fp16, y = value_layer_37_cast_fp16)[name = string("context_layer_73_cast_fp16")];
-            tensor<int32, [4]> var_4292 = const()[name = string("op_4292"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4293_cast_fp16 = reshape(shape = var_4292, x = context_layer_73_cast_fp16)[name = string("op_4293_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_73_cast_fp16 = matmul(transpose_x = context_layer_73_transpose_x_0, transpose_y = context_layer_73_transpose_y_0, x = var_4288_cast_fp16, y = value_layer_37_cast_fp16)[name = string("context_layer_73_cast_fp16")];
+            tensor<int32, [4]> var_4292 = const()[name = string("op_4292"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4293_cast_fp16 = reshape(shape = var_4292, x = context_layer_73_cast_fp16)[name = string("op_4293_cast_fp16")];
             tensor<int32, [4]> var_4294 = const()[name = string("op_4294"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_4295_cast_fp16 = transpose(perm = var_4294, x = var_4293_cast_fp16)[name = string("transpose_30")];
-            tensor<fp16, [1, 256, 1024]> input_445_cast_fp16 = reshape(shape = var_4299, x = var_4295_cast_fp16)[name = string("input_445_cast_fp16")];
+            tensor<int32, [3]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_4295_cast_fp16 = transpose(perm = var_4294, x = var_4293_cast_fp16)[name = string("transpose_30")];
+            tensor<fp16, [1, 192, 1024]> input_445_cast_fp16 = reshape(shape = var_4299, x = var_4295_cast_fp16)[name = string("input_445_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762848960)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764946176)))];
-            tensor<fp16, [1, 256, 1024]> linear_149_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_attention_output_dense_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_149_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_449_cast_fp16 = add(x = linear_149_cast_fp16, y = input_437_cast_fp16)[name = string("input_449_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_149_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_attention_output_dense_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_449_cast_fp16 = add(x = linear_149_cast_fp16, y = input_437_cast_fp16)[name = string("input_449_cast_fp16")];
             tensor<int32, [1]> input_451_axes_0 = const()[name = string("input_451_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764948288)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764950400)))];
-            tensor<fp16, [1, 256, 1024]> input_451_cast_fp16 = layer_norm(axes = input_451_axes_0, beta = encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16, x = input_449_cast_fp16)[name = string("input_451_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_451_cast_fp16 = layer_norm(axes = input_451_axes_0, beta = encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16, x = input_449_cast_fp16)[name = string("input_451_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_18_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764952512)))];
             tensor<fp16, [4096]> encoder_encoder_layer_18_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773341184)))];
-            tensor<fp16, [1, 256, 4096]> linear_150_cast_fp16 = linear(bias = encoder_encoder_layer_18_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_18_intermediate_dense_weight_to_fp16, x = input_451_cast_fp16)[name = string("linear_150_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_150_cast_fp16 = linear(bias = encoder_encoder_layer_18_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_18_intermediate_dense_weight_to_fp16, x = input_451_cast_fp16)[name = string("linear_150_cast_fp16")];
             string input_455_mode_0 = const()[name = string("input_455_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_455_cast_fp16 = gelu(mode = input_455_mode_0, x = linear_150_cast_fp16)[name = string("input_455_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_455_cast_fp16 = gelu(mode = input_455_mode_0, x = linear_150_cast_fp16)[name = string("input_455_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_18_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773349440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781738112)))];
-            tensor<fp16, [1, 256, 1024]> linear_151_cast_fp16 = linear(bias = encoder_encoder_layer_18_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_output_dense_weight_to_fp16, x = input_455_cast_fp16)[name = string("linear_151_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_459_cast_fp16 = add(x = linear_151_cast_fp16, y = input_451_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_151_cast_fp16 = linear(bias = encoder_encoder_layer_18_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_output_dense_weight_to_fp16, x = input_455_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_459_cast_fp16 = add(x = linear_151_cast_fp16, y = input_451_cast_fp16)[name = string("input_459_cast_fp16")];
             tensor<int32, [1]> input_461_axes_0 = const()[name = string("input_461_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781740224)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781742336)))];
-            tensor<fp16, [1, 256, 1024]> input_461_cast_fp16 = layer_norm(axes = input_461_axes_0, beta = encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_461_cast_fp16 = layer_norm(axes = input_461_axes_0, beta = encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781744448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783841664)))];
-            tensor<fp16, [1, 256, 1024]> linear_152_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_152_cast_fp16")];
-            tensor<int32, [4]> var_4350 = const()[name = string("op_4350"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_383_cast_fp16 = reshape(shape = var_4350, x = linear_152_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_152_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<int32, [4]> var_4350 = const()[name = string("op_4350"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_383_cast_fp16 = reshape(shape = var_4350, x = linear_152_cast_fp16)[name = string("x_383_cast_fp16")];
             tensor<int32, [4]> var_4352 = const()[name = string("op_4352"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4353_cast_fp16 = transpose(perm = var_4352, x = x_383_cast_fp16)[name = string("transpose_29")];
-            tensor<fp16, [16, 256, 64]> query_layer_117_cast_fp16 = reshape(shape = var_4357, x = var_4353_cast_fp16)[name = string("query_layer_117_cast_fp16")];
+            tensor<int32, [3]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4353_cast_fp16 = transpose(perm = var_4352, x = x_383_cast_fp16)[name = string("transpose_29")];
+            tensor<fp16, [16, 192, 64]> query_layer_117_cast_fp16 = reshape(shape = var_4357, x = var_4353_cast_fp16)[name = string("query_layer_117_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783843776)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(785940992)))];
-            tensor<fp16, [1, 256, 1024]> linear_153_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_153_cast_fp16")];
-            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_387_cast_fp16 = reshape(shape = var_4364, x = linear_153_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_153_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_153_cast_fp16")];
+            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_387_cast_fp16 = reshape(shape = var_4364, x = linear_153_cast_fp16)[name = string("x_387_cast_fp16")];
             tensor<int32, [4]> var_4366 = const()[name = string("op_4366"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4371 = const()[name = string("op_4371"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4367_cast_fp16 = transpose(perm = var_4366, x = x_387_cast_fp16)[name = string("transpose_28")];
-            tensor<fp16, [16, 256, 64]> key_layer_39_cast_fp16 = reshape(shape = var_4371, x = var_4367_cast_fp16)[name = string("key_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_4371 = const()[name = string("op_4371"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4367_cast_fp16 = transpose(perm = var_4366, x = x_387_cast_fp16)[name = string("transpose_28")];
+            tensor<fp16, [16, 192, 64]> key_layer_39_cast_fp16 = reshape(shape = var_4371, x = var_4367_cast_fp16)[name = string("key_layer_39_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(785943104)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788040320)))];
-            tensor<fp16, [1, 256, 1024]> linear_154_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_154_cast_fp16")];
-            tensor<int32, [4]> var_4378 = const()[name = string("op_4378"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_391_cast_fp16 = reshape(shape = var_4378, x = linear_154_cast_fp16)[name = string("x_391_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_154_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<int32, [4]> var_4378 = const()[name = string("op_4378"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_391_cast_fp16 = reshape(shape = var_4378, x = linear_154_cast_fp16)[name = string("x_391_cast_fp16")];
             tensor<int32, [4]> var_4380 = const()[name = string("op_4380"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4385 = const()[name = string("op_4385"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4381_cast_fp16 = transpose(perm = var_4380, x = x_391_cast_fp16)[name = string("transpose_27")];
-            tensor<fp16, [16, 256, 64]> value_layer_39_cast_fp16 = reshape(shape = var_4385, x = var_4381_cast_fp16)[name = string("value_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_4385 = const()[name = string("op_4385"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4381_cast_fp16 = transpose(perm = var_4380, x = x_391_cast_fp16)[name = string("transpose_27")];
+            tensor<fp16, [16, 192, 64]> value_layer_39_cast_fp16 = reshape(shape = var_4385, x = var_4381_cast_fp16)[name = string("value_layer_39_cast_fp16")];
             tensor<int32, [3]> var_4395_perm_0 = const()[name = string("op_4395_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4397_y_0_to_fp16 = const()[name = string("_inversed_4397_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_4395_cast_fp16 = transpose(perm = var_4395_perm_0, x = key_layer_39_cast_fp16)[name = string("transpose_26")];
-            tensor<fp16, [16, 64, 256]> _inversed_4397_cast_fp16 = mul(x = var_4395_cast_fp16, y = _inversed_4397_y_0_to_fp16)[name = string("_inversed_4397_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_4395_cast_fp16 = transpose(perm = var_4395_perm_0, x = key_layer_39_cast_fp16)[name = string("transpose_26")];
+            tensor<fp16, [16, 64, 192]> _inversed_4397_cast_fp16 = mul(x = var_4395_cast_fp16, y = _inversed_4397_y_0_to_fp16)[name = string("_inversed_4397_cast_fp16")];
             bool attention_scores_115_transpose_x_0 = const()[name = string("attention_scores_115_transpose_x_0"), val = bool(false)];
             bool attention_scores_115_transpose_y_0 = const()[name = string("attention_scores_115_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_115_cast_fp16 = matmul(transpose_x = attention_scores_115_transpose_x_0, transpose_y = attention_scores_115_transpose_y_0, x = query_layer_117_cast_fp16, y = _inversed_4397_cast_fp16)[name = string("attention_scores_115_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_115_cast_fp16 = matmul(transpose_x = attention_scores_115_transpose_x_0, transpose_y = attention_scores_115_transpose_y_0, x = query_layer_117_cast_fp16, y = _inversed_4397_cast_fp16)[name = string("attention_scores_115_cast_fp16")];
             bool c2p_att_77_transpose_x_0 = const()[name = string("c2p_att_77_transpose_x_0"), val = bool(false)];
             bool c2p_att_77_transpose_y_0 = const()[name = string("c2p_att_77_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4453_to_fp16 = const()[name = string("op_4453_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788042432)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_77_cast_fp16 = matmul(transpose_x = c2p_att_77_transpose_x_0, transpose_y = c2p_att_77_transpose_y_0, x = query_layer_117_cast_fp16, y = var_4453_to_fp16)[name = string("c2p_att_77_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_77_cast_fp16 = matmul(transpose_x = c2p_att_77_transpose_x_0, transpose_y = c2p_att_77_transpose_y_0, x = query_layer_117_cast_fp16, y = var_4453_to_fp16)[name = string("c2p_att_77_cast_fp16")];
             bool c2p_att_79_validate_indices_0 = const()[name = string("c2p_att_79_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_79_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_79_validate_indices_0, x = c2p_att_77_cast_fp16)[name = string("c2p_att_79_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_79_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_79_validate_indices_0, x = c2p_att_77_cast_fp16)[name = string("c2p_att_79_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4466_y_0_to_fp16 = const()[name = string("_inversed_4466_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_4466_cast_fp16 = mul(x = c2p_att_79_cast_fp16_cast_int16, y = _inversed_4466_y_0_to_fp16)[name = string("_inversed_4466_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_4466_cast_fp16 = mul(x = c2p_att_79_cast_fp16_cast_int16, y = _inversed_4466_y_0_to_fp16)[name = string("_inversed_4466_cast_fp16")];
             bool p2c_att_77_transpose_x_0 = const()[name = string("p2c_att_77_transpose_x_0"), val = bool(false)];
             bool p2c_att_77_transpose_y_0 = const()[name = string("p2c_att_77_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4481_to_fp16 = const()[name = string("op_4481_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789091072)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_77_cast_fp16 = matmul(transpose_x = p2c_att_77_transpose_x_0, transpose_y = p2c_att_77_transpose_y_0, x = key_layer_39_cast_fp16, y = var_4481_to_fp16)[name = string("p2c_att_77_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_77_cast_fp16 = matmul(transpose_x = p2c_att_77_transpose_x_0, transpose_y = p2c_att_77_transpose_y_0, x = key_layer_39_cast_fp16, y = var_4481_to_fp16)[name = string("p2c_att_77_cast_fp16")];
             bool var_4489_validate_indices_0 = const()[name = string("op_4489_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_4489_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4489_validate_indices_0, x = p2c_att_77_cast_fp16)[name = string("op_4489_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_4489_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4489_validate_indices_0, x = p2c_att_77_cast_fp16)[name = string("op_4489_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_79_perm_0 = const()[name = string("p2c_att_79_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4492_y_0_to_fp16 = const()[name = string("_inversed_4492_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_79_cast_fp16 = transpose(perm = p2c_att_79_perm_0, x = var_4489_cast_fp16_cast_int16)[name = string("transpose_25")];
-            tensor<fp16, [16, 256, 256]> _inversed_4492_cast_fp16 = mul(x = p2c_att_79_cast_fp16, y = _inversed_4492_y_0_to_fp16)[name = string("_inversed_4492_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_39_cast_fp16 = add(x = _inversed_4466_cast_fp16, y = _inversed_4492_cast_fp16)[name = string("rel_att_39_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_117_cast_fp16 = add(x = attention_scores_115_cast_fp16, y = rel_att_39_cast_fp16)[name = string("attention_scores_117_cast_fp16")];
-            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_119_cast_fp16 = reshape(shape = var_4497, x = attention_scores_117_cast_fp16)[name = string("attention_scores_119_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_465_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_119_cast_fp16, cond = var_339)[name = string("input_465_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_467_cast_fp16 = softmax(axis = var_56, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
-            tensor<int32, [3]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_4507_cast_fp16 = reshape(shape = var_4506, x = input_467_cast_fp16)[name = string("op_4507_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_79_cast_fp16 = transpose(perm = p2c_att_79_perm_0, x = var_4489_cast_fp16_cast_int16)[name = string("transpose_25")];
+            tensor<fp16, [16, 192, 192]> _inversed_4492_cast_fp16 = mul(x = p2c_att_79_cast_fp16, y = _inversed_4492_y_0_to_fp16)[name = string("_inversed_4492_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_39_cast_fp16 = add(x = _inversed_4466_cast_fp16, y = _inversed_4492_cast_fp16)[name = string("rel_att_39_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_117_cast_fp16 = add(x = attention_scores_115_cast_fp16, y = rel_att_39_cast_fp16)[name = string("attention_scores_117_cast_fp16")];
+            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_119_cast_fp16 = reshape(shape = var_4497, x = attention_scores_117_cast_fp16)[name = string("attention_scores_119_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_465_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_119_cast_fp16, cond = var_339)[name = string("input_465_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_467_cast_fp16 = softmax(axis = var_56, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            tensor<int32, [3]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_4507_cast_fp16 = reshape(shape = var_4506, x = input_467_cast_fp16)[name = string("op_4507_cast_fp16")];
             bool context_layer_77_transpose_x_0 = const()[name = string("context_layer_77_transpose_x_0"), val = bool(false)];
             bool context_layer_77_transpose_y_0 = const()[name = string("context_layer_77_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_77_cast_fp16 = matmul(transpose_x = context_layer_77_transpose_x_0, transpose_y = context_layer_77_transpose_y_0, x = var_4507_cast_fp16, y = value_layer_39_cast_fp16)[name = string("context_layer_77_cast_fp16")];
-            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4512_cast_fp16 = reshape(shape = var_4511, x = context_layer_77_cast_fp16)[name = string("op_4512_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_77_cast_fp16 = matmul(transpose_x = context_layer_77_transpose_x_0, transpose_y = context_layer_77_transpose_y_0, x = var_4507_cast_fp16, y = value_layer_39_cast_fp16)[name = string("context_layer_77_cast_fp16")];
+            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4512_cast_fp16 = reshape(shape = var_4511, x = context_layer_77_cast_fp16)[name = string("op_4512_cast_fp16")];
             tensor<int32, [4]> var_4513 = const()[name = string("op_4513"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4518 = const()[name = string("op_4518"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_4514_cast_fp16 = transpose(perm = var_4513, x = var_4512_cast_fp16)[name = string("transpose_24")];
-            tensor<fp16, [1, 256, 1024]> input_469_cast_fp16 = reshape(shape = var_4518, x = var_4514_cast_fp16)[name = string("input_469_cast_fp16")];
+            tensor<int32, [3]> var_4518 = const()[name = string("op_4518"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_4514_cast_fp16 = transpose(perm = var_4513, x = var_4512_cast_fp16)[name = string("transpose_24")];
+            tensor<fp16, [1, 192, 1024]> input_469_cast_fp16 = reshape(shape = var_4518, x = var_4514_cast_fp16)[name = string("input_469_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790139712)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792236928)))];
-            tensor<fp16, [1, 256, 1024]> linear_157_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_attention_output_dense_weight_to_fp16, x = input_469_cast_fp16)[name = string("linear_157_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_473_cast_fp16 = add(x = linear_157_cast_fp16, y = input_461_cast_fp16)[name = string("input_473_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_157_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_attention_output_dense_weight_to_fp16, x = input_469_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_473_cast_fp16 = add(x = linear_157_cast_fp16, y = input_461_cast_fp16)[name = string("input_473_cast_fp16")];
             tensor<int32, [1]> input_475_axes_0 = const()[name = string("input_475_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792239040)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792241152)))];
-            tensor<fp16, [1, 256, 1024]> input_475_cast_fp16 = layer_norm(axes = input_475_axes_0, beta = encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16, x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_475_cast_fp16 = layer_norm(axes = input_475_axes_0, beta = encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16, x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_19_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792243264)))];
             tensor<fp16, [4096]> encoder_encoder_layer_19_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800631936)))];
-            tensor<fp16, [1, 256, 4096]> linear_158_cast_fp16 = linear(bias = encoder_encoder_layer_19_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_19_intermediate_dense_weight_to_fp16, x = input_475_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_158_cast_fp16 = linear(bias = encoder_encoder_layer_19_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_19_intermediate_dense_weight_to_fp16, x = input_475_cast_fp16)[name = string("linear_158_cast_fp16")];
             string input_479_mode_0 = const()[name = string("input_479_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_479_cast_fp16 = gelu(mode = input_479_mode_0, x = linear_158_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_479_cast_fp16 = gelu(mode = input_479_mode_0, x = linear_158_cast_fp16)[name = string("input_479_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_19_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800640192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809028864)))];
-            tensor<fp16, [1, 256, 1024]> linear_159_cast_fp16 = linear(bias = encoder_encoder_layer_19_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_output_dense_weight_to_fp16, x = input_479_cast_fp16)[name = string("linear_159_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_483_cast_fp16 = add(x = linear_159_cast_fp16, y = input_475_cast_fp16)[name = string("input_483_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_159_cast_fp16 = linear(bias = encoder_encoder_layer_19_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_output_dense_weight_to_fp16, x = input_479_cast_fp16)[name = string("linear_159_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_483_cast_fp16 = add(x = linear_159_cast_fp16, y = input_475_cast_fp16)[name = string("input_483_cast_fp16")];
             tensor<int32, [1]> input_485_axes_0 = const()[name = string("input_485_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809030976)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809033088)))];
-            tensor<fp16, [1, 256, 1024]> input_485_cast_fp16 = layer_norm(axes = input_485_axes_0, beta = encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16, x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_485_cast_fp16 = layer_norm(axes = input_485_axes_0, beta = encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16, x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809035200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811132416)))];
-            tensor<fp16, [1, 256, 1024]> linear_160_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_160_cast_fp16")];
-            tensor<int32, [4]> var_4569 = const()[name = string("op_4569"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_403_cast_fp16 = reshape(shape = var_4569, x = linear_160_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_160_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<int32, [4]> var_4569 = const()[name = string("op_4569"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_403_cast_fp16 = reshape(shape = var_4569, x = linear_160_cast_fp16)[name = string("x_403_cast_fp16")];
             tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4572_cast_fp16 = transpose(perm = var_4571, x = x_403_cast_fp16)[name = string("transpose_23")];
-            tensor<fp16, [16, 256, 64]> query_layer_123_cast_fp16 = reshape(shape = var_4576, x = var_4572_cast_fp16)[name = string("query_layer_123_cast_fp16")];
+            tensor<int32, [3]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4572_cast_fp16 = transpose(perm = var_4571, x = x_403_cast_fp16)[name = string("transpose_23")];
+            tensor<fp16, [16, 192, 64]> query_layer_123_cast_fp16 = reshape(shape = var_4576, x = var_4572_cast_fp16)[name = string("query_layer_123_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811134528)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813231744)))];
-            tensor<fp16, [1, 256, 1024]> linear_161_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_161_cast_fp16")];
-            tensor<int32, [4]> var_4583 = const()[name = string("op_4583"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_407_cast_fp16 = reshape(shape = var_4583, x = linear_161_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_161_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<int32, [4]> var_4583 = const()[name = string("op_4583"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_407_cast_fp16 = reshape(shape = var_4583, x = linear_161_cast_fp16)[name = string("x_407_cast_fp16")];
             tensor<int32, [4]> var_4585 = const()[name = string("op_4585"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4586_cast_fp16 = transpose(perm = var_4585, x = x_407_cast_fp16)[name = string("transpose_22")];
-            tensor<fp16, [16, 256, 64]> key_layer_41_cast_fp16 = reshape(shape = var_4590, x = var_4586_cast_fp16)[name = string("key_layer_41_cast_fp16")];
+            tensor<int32, [3]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4586_cast_fp16 = transpose(perm = var_4585, x = x_407_cast_fp16)[name = string("transpose_22")];
+            tensor<fp16, [16, 192, 64]> key_layer_41_cast_fp16 = reshape(shape = var_4590, x = var_4586_cast_fp16)[name = string("key_layer_41_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813233856)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815331072)))];
-            tensor<fp16, [1, 256, 1024]> linear_162_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_162_cast_fp16")];
-            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_411_cast_fp16 = reshape(shape = var_4597, x = linear_162_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_162_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_162_cast_fp16")];
+            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_411_cast_fp16 = reshape(shape = var_4597, x = linear_162_cast_fp16)[name = string("x_411_cast_fp16")];
             tensor<int32, [4]> var_4599 = const()[name = string("op_4599"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4600_cast_fp16 = transpose(perm = var_4599, x = x_411_cast_fp16)[name = string("transpose_21")];
-            tensor<fp16, [16, 256, 64]> value_layer_41_cast_fp16 = reshape(shape = var_4604, x = var_4600_cast_fp16)[name = string("value_layer_41_cast_fp16")];
+            tensor<int32, [3]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4600_cast_fp16 = transpose(perm = var_4599, x = x_411_cast_fp16)[name = string("transpose_21")];
+            tensor<fp16, [16, 192, 64]> value_layer_41_cast_fp16 = reshape(shape = var_4604, x = var_4600_cast_fp16)[name = string("value_layer_41_cast_fp16")];
             tensor<int32, [3]> var_4614_perm_0 = const()[name = string("op_4614_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4616_y_0_to_fp16 = const()[name = string("_inversed_4616_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_4614_cast_fp16 = transpose(perm = var_4614_perm_0, x = key_layer_41_cast_fp16)[name = string("transpose_20")];
-            tensor<fp16, [16, 64, 256]> _inversed_4616_cast_fp16 = mul(x = var_4614_cast_fp16, y = _inversed_4616_y_0_to_fp16)[name = string("_inversed_4616_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_4614_cast_fp16 = transpose(perm = var_4614_perm_0, x = key_layer_41_cast_fp16)[name = string("transpose_20")];
+            tensor<fp16, [16, 64, 192]> _inversed_4616_cast_fp16 = mul(x = var_4614_cast_fp16, y = _inversed_4616_y_0_to_fp16)[name = string("_inversed_4616_cast_fp16")];
             bool attention_scores_121_transpose_x_0 = const()[name = string("attention_scores_121_transpose_x_0"), val = bool(false)];
             bool attention_scores_121_transpose_y_0 = const()[name = string("attention_scores_121_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_121_cast_fp16 = matmul(transpose_x = attention_scores_121_transpose_x_0, transpose_y = attention_scores_121_transpose_y_0, x = query_layer_123_cast_fp16, y = _inversed_4616_cast_fp16)[name = string("attention_scores_121_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_121_cast_fp16 = matmul(transpose_x = attention_scores_121_transpose_x_0, transpose_y = attention_scores_121_transpose_y_0, x = query_layer_123_cast_fp16, y = _inversed_4616_cast_fp16)[name = string("attention_scores_121_cast_fp16")];
             bool c2p_att_81_transpose_x_0 = const()[name = string("c2p_att_81_transpose_x_0"), val = bool(false)];
             bool c2p_att_81_transpose_y_0 = const()[name = string("c2p_att_81_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4672_to_fp16 = const()[name = string("op_4672_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815333184)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_81_cast_fp16 = matmul(transpose_x = c2p_att_81_transpose_x_0, transpose_y = c2p_att_81_transpose_y_0, x = query_layer_123_cast_fp16, y = var_4672_to_fp16)[name = string("c2p_att_81_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_81_cast_fp16 = matmul(transpose_x = c2p_att_81_transpose_x_0, transpose_y = c2p_att_81_transpose_y_0, x = query_layer_123_cast_fp16, y = var_4672_to_fp16)[name = string("c2p_att_81_cast_fp16")];
             bool c2p_att_83_validate_indices_0 = const()[name = string("c2p_att_83_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_83_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_83_validate_indices_0, x = c2p_att_81_cast_fp16)[name = string("c2p_att_83_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_83_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_83_validate_indices_0, x = c2p_att_81_cast_fp16)[name = string("c2p_att_83_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4685_y_0_to_fp16 = const()[name = string("_inversed_4685_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_4685_cast_fp16 = mul(x = c2p_att_83_cast_fp16_cast_int16, y = _inversed_4685_y_0_to_fp16)[name = string("_inversed_4685_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_4685_cast_fp16 = mul(x = c2p_att_83_cast_fp16_cast_int16, y = _inversed_4685_y_0_to_fp16)[name = string("_inversed_4685_cast_fp16")];
             bool p2c_att_81_transpose_x_0 = const()[name = string("p2c_att_81_transpose_x_0"), val = bool(false)];
             bool p2c_att_81_transpose_y_0 = const()[name = string("p2c_att_81_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4700_to_fp16 = const()[name = string("op_4700_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816381824)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_81_cast_fp16 = matmul(transpose_x = p2c_att_81_transpose_x_0, transpose_y = p2c_att_81_transpose_y_0, x = key_layer_41_cast_fp16, y = var_4700_to_fp16)[name = string("p2c_att_81_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_81_cast_fp16 = matmul(transpose_x = p2c_att_81_transpose_x_0, transpose_y = p2c_att_81_transpose_y_0, x = key_layer_41_cast_fp16, y = var_4700_to_fp16)[name = string("p2c_att_81_cast_fp16")];
             bool var_4708_validate_indices_0 = const()[name = string("op_4708_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_4708_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4708_validate_indices_0, x = p2c_att_81_cast_fp16)[name = string("op_4708_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_4708_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4708_validate_indices_0, x = p2c_att_81_cast_fp16)[name = string("op_4708_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_83_perm_0 = const()[name = string("p2c_att_83_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4711_y_0_to_fp16 = const()[name = string("_inversed_4711_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_83_cast_fp16 = transpose(perm = p2c_att_83_perm_0, x = var_4708_cast_fp16_cast_int16)[name = string("transpose_19")];
-            tensor<fp16, [16, 256, 256]> _inversed_4711_cast_fp16 = mul(x = p2c_att_83_cast_fp16, y = _inversed_4711_y_0_to_fp16)[name = string("_inversed_4711_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_41_cast_fp16 = add(x = _inversed_4685_cast_fp16, y = _inversed_4711_cast_fp16)[name = string("rel_att_41_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_123_cast_fp16 = add(x = attention_scores_121_cast_fp16, y = rel_att_41_cast_fp16)[name = string("attention_scores_123_cast_fp16")];
-            tensor<int32, [4]> var_4716 = const()[name = string("op_4716"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_125_cast_fp16 = reshape(shape = var_4716, x = attention_scores_123_cast_fp16)[name = string("attention_scores_125_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_489_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_125_cast_fp16, cond = var_339)[name = string("input_489_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_491_cast_fp16 = softmax(axis = var_56, x = input_489_cast_fp16)[name = string("input_491_cast_fp16")];
-            tensor<int32, [3]> var_4725 = const()[name = string("op_4725"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_4726_cast_fp16 = reshape(shape = var_4725, x = input_491_cast_fp16)[name = string("op_4726_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_83_cast_fp16 = transpose(perm = p2c_att_83_perm_0, x = var_4708_cast_fp16_cast_int16)[name = string("transpose_19")];
+            tensor<fp16, [16, 192, 192]> _inversed_4711_cast_fp16 = mul(x = p2c_att_83_cast_fp16, y = _inversed_4711_y_0_to_fp16)[name = string("_inversed_4711_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_41_cast_fp16 = add(x = _inversed_4685_cast_fp16, y = _inversed_4711_cast_fp16)[name = string("rel_att_41_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_123_cast_fp16 = add(x = attention_scores_121_cast_fp16, y = rel_att_41_cast_fp16)[name = string("attention_scores_123_cast_fp16")];
+            tensor<int32, [4]> var_4716 = const()[name = string("op_4716"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_125_cast_fp16 = reshape(shape = var_4716, x = attention_scores_123_cast_fp16)[name = string("attention_scores_125_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_489_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_125_cast_fp16, cond = var_339)[name = string("input_489_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_491_cast_fp16 = softmax(axis = var_56, x = input_489_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [3]> var_4725 = const()[name = string("op_4725"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_4726_cast_fp16 = reshape(shape = var_4725, x = input_491_cast_fp16)[name = string("op_4726_cast_fp16")];
             bool context_layer_81_transpose_x_0 = const()[name = string("context_layer_81_transpose_x_0"), val = bool(false)];
             bool context_layer_81_transpose_y_0 = const()[name = string("context_layer_81_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_81_cast_fp16 = matmul(transpose_x = context_layer_81_transpose_x_0, transpose_y = context_layer_81_transpose_y_0, x = var_4726_cast_fp16, y = value_layer_41_cast_fp16)[name = string("context_layer_81_cast_fp16")];
-            tensor<int32, [4]> var_4730 = const()[name = string("op_4730"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4731_cast_fp16 = reshape(shape = var_4730, x = context_layer_81_cast_fp16)[name = string("op_4731_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_81_cast_fp16 = matmul(transpose_x = context_layer_81_transpose_x_0, transpose_y = context_layer_81_transpose_y_0, x = var_4726_cast_fp16, y = value_layer_41_cast_fp16)[name = string("context_layer_81_cast_fp16")];
+            tensor<int32, [4]> var_4730 = const()[name = string("op_4730"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4731_cast_fp16 = reshape(shape = var_4730, x = context_layer_81_cast_fp16)[name = string("op_4731_cast_fp16")];
             tensor<int32, [4]> var_4732 = const()[name = string("op_4732"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4737 = const()[name = string("op_4737"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_4733_cast_fp16 = transpose(perm = var_4732, x = var_4731_cast_fp16)[name = string("transpose_18")];
-            tensor<fp16, [1, 256, 1024]> input_493_cast_fp16 = reshape(shape = var_4737, x = var_4733_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [3]> var_4737 = const()[name = string("op_4737"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_4733_cast_fp16 = transpose(perm = var_4732, x = var_4731_cast_fp16)[name = string("transpose_18")];
+            tensor<fp16, [1, 192, 1024]> input_493_cast_fp16 = reshape(shape = var_4737, x = var_4733_cast_fp16)[name = string("input_493_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817430464)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819527680)))];
-            tensor<fp16, [1, 256, 1024]> linear_165_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_attention_output_dense_weight_to_fp16, x = input_493_cast_fp16)[name = string("linear_165_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_497_cast_fp16 = add(x = linear_165_cast_fp16, y = input_485_cast_fp16)[name = string("input_497_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_165_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_attention_output_dense_weight_to_fp16, x = input_493_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_497_cast_fp16 = add(x = linear_165_cast_fp16, y = input_485_cast_fp16)[name = string("input_497_cast_fp16")];
             tensor<int32, [1]> input_499_axes_0 = const()[name = string("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819529792)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819531904)))];
-            tensor<fp16, [1, 256, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, beta = encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, beta = encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_20_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819534016)))];
             tensor<fp16, [4096]> encoder_encoder_layer_20_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827922688)))];
-            tensor<fp16, [1, 256, 4096]> linear_166_cast_fp16 = linear(bias = encoder_encoder_layer_20_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_20_intermediate_dense_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_166_cast_fp16 = linear(bias = encoder_encoder_layer_20_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_20_intermediate_dense_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_166_cast_fp16")];
             string input_503_mode_0 = const()[name = string("input_503_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_503_cast_fp16 = gelu(mode = input_503_mode_0, x = linear_166_cast_fp16)[name = string("input_503_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_503_cast_fp16 = gelu(mode = input_503_mode_0, x = linear_166_cast_fp16)[name = string("input_503_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_20_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827930944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836319616)))];
-            tensor<fp16, [1, 256, 1024]> linear_167_cast_fp16 = linear(bias = encoder_encoder_layer_20_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_output_dense_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_167_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_507_cast_fp16 = add(x = linear_167_cast_fp16, y = input_499_cast_fp16)[name = string("input_507_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_167_cast_fp16 = linear(bias = encoder_encoder_layer_20_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_output_dense_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_507_cast_fp16 = add(x = linear_167_cast_fp16, y = input_499_cast_fp16)[name = string("input_507_cast_fp16")];
             tensor<int32, [1]> input_509_axes_0 = const()[name = string("input_509_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836321728)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836323840)))];
-            tensor<fp16, [1, 256, 1024]> input_509_cast_fp16 = layer_norm(axes = input_509_axes_0, beta = encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_509_cast_fp16 = layer_norm(axes = input_509_axes_0, beta = encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836325952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(838423168)))];
-            tensor<fp16, [1, 256, 1024]> linear_168_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_168_cast_fp16")];
-            tensor<int32, [4]> var_4788 = const()[name = string("op_4788"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_423_cast_fp16 = reshape(shape = var_4788, x = linear_168_cast_fp16)[name = string("x_423_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_168_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_168_cast_fp16")];
+            tensor<int32, [4]> var_4788 = const()[name = string("op_4788"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_423_cast_fp16 = reshape(shape = var_4788, x = linear_168_cast_fp16)[name = string("x_423_cast_fp16")];
             tensor<int32, [4]> var_4790 = const()[name = string("op_4790"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4795 = const()[name = string("op_4795"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4791_cast_fp16 = transpose(perm = var_4790, x = x_423_cast_fp16)[name = string("transpose_17")];
-            tensor<fp16, [16, 256, 64]> query_layer_129_cast_fp16 = reshape(shape = var_4795, x = var_4791_cast_fp16)[name = string("query_layer_129_cast_fp16")];
+            tensor<int32, [3]> var_4795 = const()[name = string("op_4795"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4791_cast_fp16 = transpose(perm = var_4790, x = x_423_cast_fp16)[name = string("transpose_17")];
+            tensor<fp16, [16, 192, 64]> query_layer_129_cast_fp16 = reshape(shape = var_4795, x = var_4791_cast_fp16)[name = string("query_layer_129_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(838425280)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(840522496)))];
-            tensor<fp16, [1, 256, 1024]> linear_169_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_169_cast_fp16")];
-            tensor<int32, [4]> var_4802 = const()[name = string("op_4802"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_427_cast_fp16 = reshape(shape = var_4802, x = linear_169_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_169_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<int32, [4]> var_4802 = const()[name = string("op_4802"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_427_cast_fp16 = reshape(shape = var_4802, x = linear_169_cast_fp16)[name = string("x_427_cast_fp16")];
             tensor<int32, [4]> var_4804 = const()[name = string("op_4804"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4805_cast_fp16 = transpose(perm = var_4804, x = x_427_cast_fp16)[name = string("transpose_16")];
-            tensor<fp16, [16, 256, 64]> key_layer_43_cast_fp16 = reshape(shape = var_4809, x = var_4805_cast_fp16)[name = string("key_layer_43_cast_fp16")];
+            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4805_cast_fp16 = transpose(perm = var_4804, x = x_427_cast_fp16)[name = string("transpose_16")];
+            tensor<fp16, [16, 192, 64]> key_layer_43_cast_fp16 = reshape(shape = var_4809, x = var_4805_cast_fp16)[name = string("key_layer_43_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(840524608)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842621824)))];
-            tensor<fp16, [1, 256, 1024]> linear_170_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_170_cast_fp16")];
-            tensor<int32, [4]> var_4816 = const()[name = string("op_4816"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_431_cast_fp16 = reshape(shape = var_4816, x = linear_170_cast_fp16)[name = string("x_431_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_170_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<int32, [4]> var_4816 = const()[name = string("op_4816"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_431_cast_fp16 = reshape(shape = var_4816, x = linear_170_cast_fp16)[name = string("x_431_cast_fp16")];
             tensor<int32, [4]> var_4818 = const()[name = string("op_4818"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4823 = const()[name = string("op_4823"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4819_cast_fp16 = transpose(perm = var_4818, x = x_431_cast_fp16)[name = string("transpose_15")];
-            tensor<fp16, [16, 256, 64]> value_layer_43_cast_fp16 = reshape(shape = var_4823, x = var_4819_cast_fp16)[name = string("value_layer_43_cast_fp16")];
+            tensor<int32, [3]> var_4823 = const()[name = string("op_4823"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4819_cast_fp16 = transpose(perm = var_4818, x = x_431_cast_fp16)[name = string("transpose_15")];
+            tensor<fp16, [16, 192, 64]> value_layer_43_cast_fp16 = reshape(shape = var_4823, x = var_4819_cast_fp16)[name = string("value_layer_43_cast_fp16")];
             tensor<int32, [3]> var_4833_perm_0 = const()[name = string("op_4833_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4835_y_0_to_fp16 = const()[name = string("_inversed_4835_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_4833_cast_fp16 = transpose(perm = var_4833_perm_0, x = key_layer_43_cast_fp16)[name = string("transpose_14")];
-            tensor<fp16, [16, 64, 256]> _inversed_4835_cast_fp16 = mul(x = var_4833_cast_fp16, y = _inversed_4835_y_0_to_fp16)[name = string("_inversed_4835_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_4833_cast_fp16 = transpose(perm = var_4833_perm_0, x = key_layer_43_cast_fp16)[name = string("transpose_14")];
+            tensor<fp16, [16, 64, 192]> _inversed_4835_cast_fp16 = mul(x = var_4833_cast_fp16, y = _inversed_4835_y_0_to_fp16)[name = string("_inversed_4835_cast_fp16")];
             bool attention_scores_127_transpose_x_0 = const()[name = string("attention_scores_127_transpose_x_0"), val = bool(false)];
             bool attention_scores_127_transpose_y_0 = const()[name = string("attention_scores_127_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_127_cast_fp16 = matmul(transpose_x = attention_scores_127_transpose_x_0, transpose_y = attention_scores_127_transpose_y_0, x = query_layer_129_cast_fp16, y = _inversed_4835_cast_fp16)[name = string("attention_scores_127_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_127_cast_fp16 = matmul(transpose_x = attention_scores_127_transpose_x_0, transpose_y = attention_scores_127_transpose_y_0, x = query_layer_129_cast_fp16, y = _inversed_4835_cast_fp16)[name = string("attention_scores_127_cast_fp16")];
             bool c2p_att_85_transpose_x_0 = const()[name = string("c2p_att_85_transpose_x_0"), val = bool(false)];
             bool c2p_att_85_transpose_y_0 = const()[name = string("c2p_att_85_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842623936)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_85_cast_fp16 = matmul(transpose_x = c2p_att_85_transpose_x_0, transpose_y = c2p_att_85_transpose_y_0, x = query_layer_129_cast_fp16, y = var_4891_to_fp16)[name = string("c2p_att_85_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_85_cast_fp16 = matmul(transpose_x = c2p_att_85_transpose_x_0, transpose_y = c2p_att_85_transpose_y_0, x = query_layer_129_cast_fp16, y = var_4891_to_fp16)[name = string("c2p_att_85_cast_fp16")];
             bool c2p_att_87_validate_indices_0 = const()[name = string("c2p_att_87_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_87_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_87_validate_indices_0, x = c2p_att_85_cast_fp16)[name = string("c2p_att_87_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_87_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_87_validate_indices_0, x = c2p_att_85_cast_fp16)[name = string("c2p_att_87_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4904_y_0_to_fp16 = const()[name = string("_inversed_4904_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_4904_cast_fp16 = mul(x = c2p_att_87_cast_fp16_cast_int16, y = _inversed_4904_y_0_to_fp16)[name = string("_inversed_4904_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_4904_cast_fp16 = mul(x = c2p_att_87_cast_fp16_cast_int16, y = _inversed_4904_y_0_to_fp16)[name = string("_inversed_4904_cast_fp16")];
             bool p2c_att_85_transpose_x_0 = const()[name = string("p2c_att_85_transpose_x_0"), val = bool(false)];
             bool p2c_att_85_transpose_y_0 = const()[name = string("p2c_att_85_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4919_to_fp16 = const()[name = string("op_4919_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843672576)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_85_cast_fp16 = matmul(transpose_x = p2c_att_85_transpose_x_0, transpose_y = p2c_att_85_transpose_y_0, x = key_layer_43_cast_fp16, y = var_4919_to_fp16)[name = string("p2c_att_85_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_85_cast_fp16 = matmul(transpose_x = p2c_att_85_transpose_x_0, transpose_y = p2c_att_85_transpose_y_0, x = key_layer_43_cast_fp16, y = var_4919_to_fp16)[name = string("p2c_att_85_cast_fp16")];
             bool var_4927_validate_indices_0 = const()[name = string("op_4927_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_4927_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4927_validate_indices_0, x = p2c_att_85_cast_fp16)[name = string("op_4927_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_4927_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4927_validate_indices_0, x = p2c_att_85_cast_fp16)[name = string("op_4927_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_87_perm_0 = const()[name = string("p2c_att_87_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4930_y_0_to_fp16 = const()[name = string("_inversed_4930_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_87_cast_fp16 = transpose(perm = p2c_att_87_perm_0, x = var_4927_cast_fp16_cast_int16)[name = string("transpose_13")];
-            tensor<fp16, [16, 256, 256]> _inversed_4930_cast_fp16 = mul(x = p2c_att_87_cast_fp16, y = _inversed_4930_y_0_to_fp16)[name = string("_inversed_4930_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_43_cast_fp16 = add(x = _inversed_4904_cast_fp16, y = _inversed_4930_cast_fp16)[name = string("rel_att_43_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_129_cast_fp16 = add(x = attention_scores_127_cast_fp16, y = rel_att_43_cast_fp16)[name = string("attention_scores_129_cast_fp16")];
-            tensor<int32, [4]> var_4935 = const()[name = string("op_4935"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_131_cast_fp16 = reshape(shape = var_4935, x = attention_scores_129_cast_fp16)[name = string("attention_scores_131_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_513_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_131_cast_fp16, cond = var_339)[name = string("input_513_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_515_cast_fp16 = softmax(axis = var_56, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
-            tensor<int32, [3]> var_4944 = const()[name = string("op_4944"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_4945_cast_fp16 = reshape(shape = var_4944, x = input_515_cast_fp16)[name = string("op_4945_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_87_cast_fp16 = transpose(perm = p2c_att_87_perm_0, x = var_4927_cast_fp16_cast_int16)[name = string("transpose_13")];
+            tensor<fp16, [16, 192, 192]> _inversed_4930_cast_fp16 = mul(x = p2c_att_87_cast_fp16, y = _inversed_4930_y_0_to_fp16)[name = string("_inversed_4930_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_43_cast_fp16 = add(x = _inversed_4904_cast_fp16, y = _inversed_4930_cast_fp16)[name = string("rel_att_43_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_129_cast_fp16 = add(x = attention_scores_127_cast_fp16, y = rel_att_43_cast_fp16)[name = string("attention_scores_129_cast_fp16")];
+            tensor<int32, [4]> var_4935 = const()[name = string("op_4935"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_131_cast_fp16 = reshape(shape = var_4935, x = attention_scores_129_cast_fp16)[name = string("attention_scores_131_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_513_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_131_cast_fp16, cond = var_339)[name = string("input_513_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_515_cast_fp16 = softmax(axis = var_56, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [3]> var_4944 = const()[name = string("op_4944"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_4945_cast_fp16 = reshape(shape = var_4944, x = input_515_cast_fp16)[name = string("op_4945_cast_fp16")];
             bool context_layer_85_transpose_x_0 = const()[name = string("context_layer_85_transpose_x_0"), val = bool(false)];
             bool context_layer_85_transpose_y_0 = const()[name = string("context_layer_85_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_85_cast_fp16 = matmul(transpose_x = context_layer_85_transpose_x_0, transpose_y = context_layer_85_transpose_y_0, x = var_4945_cast_fp16, y = value_layer_43_cast_fp16)[name = string("context_layer_85_cast_fp16")];
-            tensor<int32, [4]> var_4949 = const()[name = string("op_4949"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_4950_cast_fp16 = reshape(shape = var_4949, x = context_layer_85_cast_fp16)[name = string("op_4950_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_85_cast_fp16 = matmul(transpose_x = context_layer_85_transpose_x_0, transpose_y = context_layer_85_transpose_y_0, x = var_4945_cast_fp16, y = value_layer_43_cast_fp16)[name = string("context_layer_85_cast_fp16")];
+            tensor<int32, [4]> var_4949 = const()[name = string("op_4949"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_4950_cast_fp16 = reshape(shape = var_4949, x = context_layer_85_cast_fp16)[name = string("op_4950_cast_fp16")];
             tensor<int32, [4]> var_4951 = const()[name = string("op_4951"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4956 = const()[name = string("op_4956"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_4952_cast_fp16 = transpose(perm = var_4951, x = var_4950_cast_fp16)[name = string("transpose_12")];
-            tensor<fp16, [1, 256, 1024]> input_517_cast_fp16 = reshape(shape = var_4956, x = var_4952_cast_fp16)[name = string("input_517_cast_fp16")];
+            tensor<int32, [3]> var_4956 = const()[name = string("op_4956"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_4952_cast_fp16 = transpose(perm = var_4951, x = var_4950_cast_fp16)[name = string("transpose_12")];
+            tensor<fp16, [1, 192, 1024]> input_517_cast_fp16 = reshape(shape = var_4956, x = var_4952_cast_fp16)[name = string("input_517_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844721216)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846818432)))];
-            tensor<fp16, [1, 256, 1024]> linear_173_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_attention_output_dense_weight_to_fp16, x = input_517_cast_fp16)[name = string("linear_173_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_521_cast_fp16 = add(x = linear_173_cast_fp16, y = input_509_cast_fp16)[name = string("input_521_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_173_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_attention_output_dense_weight_to_fp16, x = input_517_cast_fp16)[name = string("linear_173_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_521_cast_fp16 = add(x = linear_173_cast_fp16, y = input_509_cast_fp16)[name = string("input_521_cast_fp16")];
             tensor<int32, [1]> input_523_axes_0 = const()[name = string("input_523_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846820544)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846822656)))];
-            tensor<fp16, [1, 256, 1024]> input_523_cast_fp16 = layer_norm(axes = input_523_axes_0, beta = encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_523_cast_fp16 = layer_norm(axes = input_523_axes_0, beta = encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_21_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846824768)))];
             tensor<fp16, [4096]> encoder_encoder_layer_21_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855213440)))];
-            tensor<fp16, [1, 256, 4096]> linear_174_cast_fp16 = linear(bias = encoder_encoder_layer_21_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_21_intermediate_dense_weight_to_fp16, x = input_523_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_174_cast_fp16 = linear(bias = encoder_encoder_layer_21_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_21_intermediate_dense_weight_to_fp16, x = input_523_cast_fp16)[name = string("linear_174_cast_fp16")];
             string input_527_mode_0 = const()[name = string("input_527_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_527_cast_fp16 = gelu(mode = input_527_mode_0, x = linear_174_cast_fp16)[name = string("input_527_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_527_cast_fp16 = gelu(mode = input_527_mode_0, x = linear_174_cast_fp16)[name = string("input_527_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_21_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855221696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863610368)))];
-            tensor<fp16, [1, 256, 1024]> linear_175_cast_fp16 = linear(bias = encoder_encoder_layer_21_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_output_dense_weight_to_fp16, x = input_527_cast_fp16)[name = string("linear_175_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_531_cast_fp16 = add(x = linear_175_cast_fp16, y = input_523_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_175_cast_fp16 = linear(bias = encoder_encoder_layer_21_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_output_dense_weight_to_fp16, x = input_527_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_531_cast_fp16 = add(x = linear_175_cast_fp16, y = input_523_cast_fp16)[name = string("input_531_cast_fp16")];
             tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863612480)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863614592)))];
-            tensor<fp16, [1, 256, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863616704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(865713920)))];
-            tensor<fp16, [1, 256, 1024]> linear_176_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_176_cast_fp16")];
-            tensor<int32, [4]> var_5007 = const()[name = string("op_5007"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_443_cast_fp16 = reshape(shape = var_5007, x = linear_176_cast_fp16)[name = string("x_443_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_176_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_5007 = const()[name = string("op_5007"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_443_cast_fp16 = reshape(shape = var_5007, x = linear_176_cast_fp16)[name = string("x_443_cast_fp16")];
             tensor<int32, [4]> var_5009 = const()[name = string("op_5009"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5014 = const()[name = string("op_5014"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5010_cast_fp16 = transpose(perm = var_5009, x = x_443_cast_fp16)[name = string("transpose_11")];
-            tensor<fp16, [16, 256, 64]> query_layer_135_cast_fp16 = reshape(shape = var_5014, x = var_5010_cast_fp16)[name = string("query_layer_135_cast_fp16")];
+            tensor<int32, [3]> var_5014 = const()[name = string("op_5014"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5010_cast_fp16 = transpose(perm = var_5009, x = x_443_cast_fp16)[name = string("transpose_11")];
+            tensor<fp16, [16, 192, 64]> query_layer_135_cast_fp16 = reshape(shape = var_5014, x = var_5010_cast_fp16)[name = string("query_layer_135_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(865716032)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867813248)))];
-            tensor<fp16, [1, 256, 1024]> linear_177_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_177_cast_fp16")];
-            tensor<int32, [4]> var_5021 = const()[name = string("op_5021"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_447_cast_fp16 = reshape(shape = var_5021, x = linear_177_cast_fp16)[name = string("x_447_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_177_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_177_cast_fp16")];
+            tensor<int32, [4]> var_5021 = const()[name = string("op_5021"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_447_cast_fp16 = reshape(shape = var_5021, x = linear_177_cast_fp16)[name = string("x_447_cast_fp16")];
             tensor<int32, [4]> var_5023 = const()[name = string("op_5023"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5028 = const()[name = string("op_5028"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5024_cast_fp16 = transpose(perm = var_5023, x = x_447_cast_fp16)[name = string("transpose_10")];
-            tensor<fp16, [16, 256, 64]> key_layer_45_cast_fp16 = reshape(shape = var_5028, x = var_5024_cast_fp16)[name = string("key_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_5028 = const()[name = string("op_5028"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5024_cast_fp16 = transpose(perm = var_5023, x = x_447_cast_fp16)[name = string("transpose_10")];
+            tensor<fp16, [16, 192, 64]> key_layer_45_cast_fp16 = reshape(shape = var_5028, x = var_5024_cast_fp16)[name = string("key_layer_45_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867815360)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(869912576)))];
-            tensor<fp16, [1, 256, 1024]> linear_178_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_178_cast_fp16")];
-            tensor<int32, [4]> var_5035 = const()[name = string("op_5035"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_451_cast_fp16 = reshape(shape = var_5035, x = linear_178_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_178_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<int32, [4]> var_5035 = const()[name = string("op_5035"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_451_cast_fp16 = reshape(shape = var_5035, x = linear_178_cast_fp16)[name = string("x_451_cast_fp16")];
             tensor<int32, [4]> var_5037 = const()[name = string("op_5037"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5042 = const()[name = string("op_5042"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5038_cast_fp16 = transpose(perm = var_5037, x = x_451_cast_fp16)[name = string("transpose_9")];
-            tensor<fp16, [16, 256, 64]> value_layer_45_cast_fp16 = reshape(shape = var_5042, x = var_5038_cast_fp16)[name = string("value_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_5042 = const()[name = string("op_5042"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5038_cast_fp16 = transpose(perm = var_5037, x = x_451_cast_fp16)[name = string("transpose_9")];
+            tensor<fp16, [16, 192, 64]> value_layer_45_cast_fp16 = reshape(shape = var_5042, x = var_5038_cast_fp16)[name = string("value_layer_45_cast_fp16")];
             tensor<int32, [3]> var_5052_perm_0 = const()[name = string("op_5052_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5054_y_0_to_fp16 = const()[name = string("_inversed_5054_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_5052_cast_fp16 = transpose(perm = var_5052_perm_0, x = key_layer_45_cast_fp16)[name = string("transpose_8")];
-            tensor<fp16, [16, 64, 256]> _inversed_5054_cast_fp16 = mul(x = var_5052_cast_fp16, y = _inversed_5054_y_0_to_fp16)[name = string("_inversed_5054_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_5052_cast_fp16 = transpose(perm = var_5052_perm_0, x = key_layer_45_cast_fp16)[name = string("transpose_8")];
+            tensor<fp16, [16, 64, 192]> _inversed_5054_cast_fp16 = mul(x = var_5052_cast_fp16, y = _inversed_5054_y_0_to_fp16)[name = string("_inversed_5054_cast_fp16")];
             bool attention_scores_133_transpose_x_0 = const()[name = string("attention_scores_133_transpose_x_0"), val = bool(false)];
             bool attention_scores_133_transpose_y_0 = const()[name = string("attention_scores_133_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_133_cast_fp16 = matmul(transpose_x = attention_scores_133_transpose_x_0, transpose_y = attention_scores_133_transpose_y_0, x = query_layer_135_cast_fp16, y = _inversed_5054_cast_fp16)[name = string("attention_scores_133_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_133_cast_fp16 = matmul(transpose_x = attention_scores_133_transpose_x_0, transpose_y = attention_scores_133_transpose_y_0, x = query_layer_135_cast_fp16, y = _inversed_5054_cast_fp16)[name = string("attention_scores_133_cast_fp16")];
             bool c2p_att_89_transpose_x_0 = const()[name = string("c2p_att_89_transpose_x_0"), val = bool(false)];
             bool c2p_att_89_transpose_y_0 = const()[name = string("c2p_att_89_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5110_to_fp16 = const()[name = string("op_5110_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(869914688)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_89_cast_fp16 = matmul(transpose_x = c2p_att_89_transpose_x_0, transpose_y = c2p_att_89_transpose_y_0, x = query_layer_135_cast_fp16, y = var_5110_to_fp16)[name = string("c2p_att_89_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_89_cast_fp16 = matmul(transpose_x = c2p_att_89_transpose_x_0, transpose_y = c2p_att_89_transpose_y_0, x = query_layer_135_cast_fp16, y = var_5110_to_fp16)[name = string("c2p_att_89_cast_fp16")];
             bool c2p_att_91_validate_indices_0 = const()[name = string("c2p_att_91_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_91_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_91_validate_indices_0, x = c2p_att_89_cast_fp16)[name = string("c2p_att_91_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_91_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_91_validate_indices_0, x = c2p_att_89_cast_fp16)[name = string("c2p_att_91_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_5123_y_0_to_fp16 = const()[name = string("_inversed_5123_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_5123_cast_fp16 = mul(x = c2p_att_91_cast_fp16_cast_int16, y = _inversed_5123_y_0_to_fp16)[name = string("_inversed_5123_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_5123_cast_fp16 = mul(x = c2p_att_91_cast_fp16_cast_int16, y = _inversed_5123_y_0_to_fp16)[name = string("_inversed_5123_cast_fp16")];
             bool p2c_att_89_transpose_x_0 = const()[name = string("p2c_att_89_transpose_x_0"), val = bool(false)];
             bool p2c_att_89_transpose_y_0 = const()[name = string("p2c_att_89_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5138_to_fp16 = const()[name = string("op_5138_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(870963328)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_89_cast_fp16 = matmul(transpose_x = p2c_att_89_transpose_x_0, transpose_y = p2c_att_89_transpose_y_0, x = key_layer_45_cast_fp16, y = var_5138_to_fp16)[name = string("p2c_att_89_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_89_cast_fp16 = matmul(transpose_x = p2c_att_89_transpose_x_0, transpose_y = p2c_att_89_transpose_y_0, x = key_layer_45_cast_fp16, y = var_5138_to_fp16)[name = string("p2c_att_89_cast_fp16")];
             bool var_5146_validate_indices_0 = const()[name = string("op_5146_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_5146_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5146_validate_indices_0, x = p2c_att_89_cast_fp16)[name = string("op_5146_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_5146_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5146_validate_indices_0, x = p2c_att_89_cast_fp16)[name = string("op_5146_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_91_perm_0 = const()[name = string("p2c_att_91_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5149_y_0_to_fp16 = const()[name = string("_inversed_5149_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_91_cast_fp16 = transpose(perm = p2c_att_91_perm_0, x = var_5146_cast_fp16_cast_int16)[name = string("transpose_7")];
-            tensor<fp16, [16, 256, 256]> _inversed_5149_cast_fp16 = mul(x = p2c_att_91_cast_fp16, y = _inversed_5149_y_0_to_fp16)[name = string("_inversed_5149_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_45_cast_fp16 = add(x = _inversed_5123_cast_fp16, y = _inversed_5149_cast_fp16)[name = string("rel_att_45_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_135_cast_fp16 = add(x = attention_scores_133_cast_fp16, y = rel_att_45_cast_fp16)[name = string("attention_scores_135_cast_fp16")];
-            tensor<int32, [4]> var_5154 = const()[name = string("op_5154"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_137_cast_fp16 = reshape(shape = var_5154, x = attention_scores_135_cast_fp16)[name = string("attention_scores_137_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_537_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_137_cast_fp16, cond = var_339)[name = string("input_537_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_539_cast_fp16 = softmax(axis = var_56, x = input_537_cast_fp16)[name = string("input_539_cast_fp16")];
-            tensor<int32, [3]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_5164_cast_fp16 = reshape(shape = var_5163, x = input_539_cast_fp16)[name = string("op_5164_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_91_cast_fp16 = transpose(perm = p2c_att_91_perm_0, x = var_5146_cast_fp16_cast_int16)[name = string("transpose_7")];
+            tensor<fp16, [16, 192, 192]> _inversed_5149_cast_fp16 = mul(x = p2c_att_91_cast_fp16, y = _inversed_5149_y_0_to_fp16)[name = string("_inversed_5149_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_45_cast_fp16 = add(x = _inversed_5123_cast_fp16, y = _inversed_5149_cast_fp16)[name = string("rel_att_45_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_135_cast_fp16 = add(x = attention_scores_133_cast_fp16, y = rel_att_45_cast_fp16)[name = string("attention_scores_135_cast_fp16")];
+            tensor<int32, [4]> var_5154 = const()[name = string("op_5154"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_137_cast_fp16 = reshape(shape = var_5154, x = attention_scores_135_cast_fp16)[name = string("attention_scores_137_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_537_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_137_cast_fp16, cond = var_339)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_539_cast_fp16 = softmax(axis = var_56, x = input_537_cast_fp16)[name = string("input_539_cast_fp16")];
+            tensor<int32, [3]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_5164_cast_fp16 = reshape(shape = var_5163, x = input_539_cast_fp16)[name = string("op_5164_cast_fp16")];
             bool context_layer_89_transpose_x_0 = const()[name = string("context_layer_89_transpose_x_0"), val = bool(false)];
             bool context_layer_89_transpose_y_0 = const()[name = string("context_layer_89_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_89_cast_fp16 = matmul(transpose_x = context_layer_89_transpose_x_0, transpose_y = context_layer_89_transpose_y_0, x = var_5164_cast_fp16, y = value_layer_45_cast_fp16)[name = string("context_layer_89_cast_fp16")];
-            tensor<int32, [4]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5169_cast_fp16 = reshape(shape = var_5168, x = context_layer_89_cast_fp16)[name = string("op_5169_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_89_cast_fp16 = matmul(transpose_x = context_layer_89_transpose_x_0, transpose_y = context_layer_89_transpose_y_0, x = var_5164_cast_fp16, y = value_layer_45_cast_fp16)[name = string("context_layer_89_cast_fp16")];
+            tensor<int32, [4]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5169_cast_fp16 = reshape(shape = var_5168, x = context_layer_89_cast_fp16)[name = string("op_5169_cast_fp16")];
             tensor<int32, [4]> var_5170 = const()[name = string("op_5170"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_5171_cast_fp16 = transpose(perm = var_5170, x = var_5169_cast_fp16)[name = string("transpose_6")];
-            tensor<fp16, [1, 256, 1024]> input_541_cast_fp16 = reshape(shape = var_5175, x = var_5171_cast_fp16)[name = string("input_541_cast_fp16")];
+            tensor<int32, [3]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_5171_cast_fp16 = transpose(perm = var_5170, x = var_5169_cast_fp16)[name = string("transpose_6")];
+            tensor<fp16, [1, 192, 1024]> input_541_cast_fp16 = reshape(shape = var_5175, x = var_5171_cast_fp16)[name = string("input_541_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872011968)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874109184)))];
-            tensor<fp16, [1, 256, 1024]> linear_181_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_attention_output_dense_weight_to_fp16, x = input_541_cast_fp16)[name = string("linear_181_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_545_cast_fp16 = add(x = linear_181_cast_fp16, y = input_533_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_181_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_attention_output_dense_weight_to_fp16, x = input_541_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_545_cast_fp16 = add(x = linear_181_cast_fp16, y = input_533_cast_fp16)[name = string("input_545_cast_fp16")];
             tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874111296)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874113408)))];
-            tensor<fp16, [1, 256, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_22_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874115520)))];
             tensor<fp16, [4096]> encoder_encoder_layer_22_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(882504192)))];
-            tensor<fp16, [1, 256, 4096]> linear_182_cast_fp16 = linear(bias = encoder_encoder_layer_22_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_22_intermediate_dense_weight_to_fp16, x = input_547_cast_fp16)[name = string("linear_182_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_182_cast_fp16 = linear(bias = encoder_encoder_layer_22_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_22_intermediate_dense_weight_to_fp16, x = input_547_cast_fp16)[name = string("linear_182_cast_fp16")];
             string input_551_mode_0 = const()[name = string("input_551_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_551_cast_fp16 = gelu(mode = input_551_mode_0, x = linear_182_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_551_cast_fp16 = gelu(mode = input_551_mode_0, x = linear_182_cast_fp16)[name = string("input_551_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_22_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(882512448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890901120)))];
-            tensor<fp16, [1, 256, 1024]> linear_183_cast_fp16 = linear(bias = encoder_encoder_layer_22_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_output_dense_weight_to_fp16, x = input_551_cast_fp16)[name = string("linear_183_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_555_cast_fp16 = add(x = linear_183_cast_fp16, y = input_547_cast_fp16)[name = string("input_555_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_183_cast_fp16 = linear(bias = encoder_encoder_layer_22_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_output_dense_weight_to_fp16, x = input_551_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_555_cast_fp16 = add(x = linear_183_cast_fp16, y = input_547_cast_fp16)[name = string("input_555_cast_fp16")];
             tensor<int32, [1]> input_557_axes_0 = const()[name = string("input_557_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890903232)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890905344)))];
-            tensor<fp16, [1, 256, 1024]> input_557_cast_fp16 = layer_norm(axes = input_557_axes_0, beta = encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16, x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_557_cast_fp16 = layer_norm(axes = input_557_axes_0, beta = encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16, x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890907456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893004672)))];
-            tensor<fp16, [1, 256, 1024]> linear_184_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_184_cast_fp16")];
-            tensor<int32, [4]> var_5226 = const()[name = string("op_5226"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_463_cast_fp16 = reshape(shape = var_5226, x = linear_184_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_184_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_5226 = const()[name = string("op_5226"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_463_cast_fp16 = reshape(shape = var_5226, x = linear_184_cast_fp16)[name = string("x_463_cast_fp16")];
             tensor<int32, [4]> var_5228 = const()[name = string("op_5228"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5229_cast_fp16 = transpose(perm = var_5228, x = x_463_cast_fp16)[name = string("transpose_5")];
-            tensor<fp16, [16, 256, 64]> query_layer_141_cast_fp16 = reshape(shape = var_5233, x = var_5229_cast_fp16)[name = string("query_layer_141_cast_fp16")];
+            tensor<int32, [3]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5229_cast_fp16 = transpose(perm = var_5228, x = x_463_cast_fp16)[name = string("transpose_5")];
+            tensor<fp16, [16, 192, 64]> query_layer_141_cast_fp16 = reshape(shape = var_5233, x = var_5229_cast_fp16)[name = string("query_layer_141_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893006784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(895104000)))];
-            tensor<fp16, [1, 256, 1024]> linear_185_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_185_cast_fp16")];
-            tensor<int32, [4]> var_5240 = const()[name = string("op_5240"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_467_cast_fp16 = reshape(shape = var_5240, x = linear_185_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_185_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_5240 = const()[name = string("op_5240"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_467_cast_fp16 = reshape(shape = var_5240, x = linear_185_cast_fp16)[name = string("x_467_cast_fp16")];
             tensor<int32, [4]> var_5242 = const()[name = string("op_5242"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5247 = const()[name = string("op_5247"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5243_cast_fp16 = transpose(perm = var_5242, x = x_467_cast_fp16)[name = string("transpose_4")];
-            tensor<fp16, [16, 256, 64]> key_layer_cast_fp16 = reshape(shape = var_5247, x = var_5243_cast_fp16)[name = string("key_layer_cast_fp16")];
+            tensor<int32, [3]> var_5247 = const()[name = string("op_5247"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5243_cast_fp16 = transpose(perm = var_5242, x = x_467_cast_fp16)[name = string("transpose_4")];
+            tensor<fp16, [16, 192, 64]> key_layer_cast_fp16 = reshape(shape = var_5247, x = var_5243_cast_fp16)[name = string("key_layer_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(895106112)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897203328)))];
-            tensor<fp16, [1, 256, 1024]> linear_186_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_186_cast_fp16")];
-            tensor<int32, [4]> var_5254 = const()[name = string("op_5254"), val = tensor<int32, [4]>([1, 256, 16, -1])];
-            tensor<fp16, [1, 256, 16, 64]> x_471_cast_fp16 = reshape(shape = var_5254, x = linear_186_cast_fp16)[name = string("x_471_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_186_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_186_cast_fp16")];
+            tensor<int32, [4]> var_5254 = const()[name = string("op_5254"), val = tensor<int32, [4]>([1, 192, 16, -1])];
+            tensor<fp16, [1, 192, 16, 64]> x_471_cast_fp16 = reshape(shape = var_5254, x = linear_186_cast_fp16)[name = string("x_471_cast_fp16")];
             tensor<int32, [4]> var_5256 = const()[name = string("op_5256"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5261 = const()[name = string("op_5261"), val = tensor<int32, [3]>([-1, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5257_cast_fp16 = transpose(perm = var_5256, x = x_471_cast_fp16)[name = string("transpose_3")];
-            tensor<fp16, [16, 256, 64]> value_layer_cast_fp16 = reshape(shape = var_5261, x = var_5257_cast_fp16)[name = string("value_layer_cast_fp16")];
+            tensor<int32, [3]> var_5261 = const()[name = string("op_5261"), val = tensor<int32, [3]>([-1, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5257_cast_fp16 = transpose(perm = var_5256, x = x_471_cast_fp16)[name = string("transpose_3")];
+            tensor<fp16, [16, 192, 64]> value_layer_cast_fp16 = reshape(shape = var_5261, x = var_5257_cast_fp16)[name = string("value_layer_cast_fp16")];
             tensor<int32, [3]> var_5271_perm_0 = const()[name = string("op_5271_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5273_y_0_to_fp16 = const()[name = string("_inversed_5273_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 256]> var_5271_cast_fp16 = transpose(perm = var_5271_perm_0, x = key_layer_cast_fp16)[name = string("transpose_2")];
-            tensor<fp16, [16, 64, 256]> _inversed_5273_cast_fp16 = mul(x = var_5271_cast_fp16, y = _inversed_5273_y_0_to_fp16)[name = string("_inversed_5273_cast_fp16")];
+            tensor<fp16, [16, 64, 192]> var_5271_cast_fp16 = transpose(perm = var_5271_perm_0, x = key_layer_cast_fp16)[name = string("transpose_2")];
+            tensor<fp16, [16, 64, 192]> _inversed_5273_cast_fp16 = mul(x = var_5271_cast_fp16, y = _inversed_5273_y_0_to_fp16)[name = string("_inversed_5273_cast_fp16")];
             bool attention_scores_139_transpose_x_0 = const()[name = string("attention_scores_139_transpose_x_0"), val = bool(false)];
             bool attention_scores_139_transpose_y_0 = const()[name = string("attention_scores_139_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> attention_scores_139_cast_fp16 = matmul(transpose_x = attention_scores_139_transpose_x_0, transpose_y = attention_scores_139_transpose_y_0, x = query_layer_141_cast_fp16, y = _inversed_5273_cast_fp16)[name = string("attention_scores_139_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_139_cast_fp16 = matmul(transpose_x = attention_scores_139_transpose_x_0, transpose_y = attention_scores_139_transpose_y_0, x = query_layer_141_cast_fp16, y = _inversed_5273_cast_fp16)[name = string("attention_scores_139_cast_fp16")];
             bool c2p_att_93_transpose_x_0 = const()[name = string("c2p_att_93_transpose_x_0"), val = bool(false)];
             bool c2p_att_93_transpose_y_0 = const()[name = string("c2p_att_93_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5329_to_fp16 = const()[name = string("op_5329_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897205440)))];
-            tensor<fp16, [16, 256, 512]> c2p_att_93_cast_fp16 = matmul(transpose_x = c2p_att_93_transpose_x_0, transpose_y = c2p_att_93_transpose_y_0, x = query_layer_141_cast_fp16, y = var_5329_to_fp16)[name = string("c2p_att_93_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> c2p_att_93_cast_fp16 = matmul(transpose_x = c2p_att_93_transpose_x_0, transpose_y = c2p_att_93_transpose_y_0, x = query_layer_141_cast_fp16, y = var_5329_to_fp16)[name = string("c2p_att_93_cast_fp16")];
             bool c2p_att_validate_indices_0 = const()[name = string("c2p_att_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> c2p_att_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_validate_indices_0, x = c2p_att_93_cast_fp16)[name = string("c2p_att_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> c2p_att_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_validate_indices_0, x = c2p_att_93_cast_fp16)[name = string("c2p_att_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_5342_y_0_to_fp16 = const()[name = string("_inversed_5342_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> _inversed_5342_cast_fp16 = mul(x = c2p_att_cast_fp16_cast_int16, y = _inversed_5342_y_0_to_fp16)[name = string("_inversed_5342_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> _inversed_5342_cast_fp16 = mul(x = c2p_att_cast_fp16_cast_int16, y = _inversed_5342_y_0_to_fp16)[name = string("_inversed_5342_cast_fp16")];
             bool p2c_att_93_transpose_x_0 = const()[name = string("p2c_att_93_transpose_x_0"), val = bool(false)];
             bool p2c_att_93_transpose_y_0 = const()[name = string("p2c_att_93_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(898254080)))];
-            tensor<fp16, [16, 256, 512]> p2c_att_93_cast_fp16 = matmul(transpose_x = p2c_att_93_transpose_x_0, transpose_y = p2c_att_93_transpose_y_0, x = key_layer_cast_fp16, y = var_5357_to_fp16)[name = string("p2c_att_93_cast_fp16")];
+            tensor<fp16, [16, 192, 512]> p2c_att_93_cast_fp16 = matmul(transpose_x = p2c_att_93_transpose_x_0, transpose_y = p2c_att_93_transpose_y_0, x = key_layer_cast_fp16, y = var_5357_to_fp16)[name = string("p2c_att_93_cast_fp16")];
             bool var_5365_validate_indices_0 = const()[name = string("op_5365_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 256]> var_5365_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5365_validate_indices_0, x = p2c_att_93_cast_fp16)[name = string("op_5365_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 192, 192]> var_5365_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5365_validate_indices_0, x = p2c_att_93_cast_fp16)[name = string("op_5365_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_perm_0 = const()[name = string("p2c_att_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5368_y_0_to_fp16 = const()[name = string("_inversed_5368_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 256, 256]> p2c_att_cast_fp16 = transpose(perm = p2c_att_perm_0, x = var_5365_cast_fp16_cast_int16)[name = string("transpose_1")];
-            tensor<fp16, [16, 256, 256]> _inversed_5368_cast_fp16 = mul(x = p2c_att_cast_fp16, y = _inversed_5368_y_0_to_fp16)[name = string("_inversed_5368_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> rel_att_cast_fp16 = add(x = _inversed_5342_cast_fp16, y = _inversed_5368_cast_fp16)[name = string("rel_att_cast_fp16")];
-            tensor<fp16, [16, 256, 256]> attention_scores_141_cast_fp16 = add(x = attention_scores_139_cast_fp16, y = rel_att_cast_fp16)[name = string("attention_scores_141_cast_fp16")];
-            tensor<int32, [4]> var_5373 = const()[name = string("op_5373"), val = tensor<int32, [4]>([-1, 16, 256, 256])];
-            tensor<fp16, [1, 16, 256, 256]> attention_scores_cast_fp16 = reshape(shape = var_5373, x = attention_scores_141_cast_fp16)[name = string("attention_scores_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_561_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_cast_fp16, cond = var_339)[name = string("input_561_cast_fp16")];
-            tensor<fp16, [1, 16, 256, 256]> input_563_cast_fp16 = softmax(axis = var_56, x = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
-            tensor<int32, [3]> var_5382 = const()[name = string("op_5382"), val = tensor<int32, [3]>([-1, 256, 256])];
-            tensor<fp16, [16, 256, 256]> var_5383_cast_fp16 = reshape(shape = var_5382, x = input_563_cast_fp16)[name = string("op_5383_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> p2c_att_cast_fp16 = transpose(perm = p2c_att_perm_0, x = var_5365_cast_fp16_cast_int16)[name = string("transpose_1")];
+            tensor<fp16, [16, 192, 192]> _inversed_5368_cast_fp16 = mul(x = p2c_att_cast_fp16, y = _inversed_5368_y_0_to_fp16)[name = string("_inversed_5368_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> rel_att_cast_fp16 = add(x = _inversed_5342_cast_fp16, y = _inversed_5368_cast_fp16)[name = string("rel_att_cast_fp16")];
+            tensor<fp16, [16, 192, 192]> attention_scores_141_cast_fp16 = add(x = attention_scores_139_cast_fp16, y = rel_att_cast_fp16)[name = string("attention_scores_141_cast_fp16")];
+            tensor<int32, [4]> var_5373 = const()[name = string("op_5373"), val = tensor<int32, [4]>([-1, 16, 192, 192])];
+            tensor<fp16, [1, 16, 192, 192]> attention_scores_cast_fp16 = reshape(shape = var_5373, x = attention_scores_141_cast_fp16)[name = string("attention_scores_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_561_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_cast_fp16, cond = var_339)[name = string("input_561_cast_fp16")];
+            tensor<fp16, [1, 16, 192, 192]> input_563_cast_fp16 = softmax(axis = var_56, x = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<int32, [3]> var_5382 = const()[name = string("op_5382"), val = tensor<int32, [3]>([-1, 192, 192])];
+            tensor<fp16, [16, 192, 192]> var_5383_cast_fp16 = reshape(shape = var_5382, x = input_563_cast_fp16)[name = string("op_5383_cast_fp16")];
             bool context_layer_93_transpose_x_0 = const()[name = string("context_layer_93_transpose_x_0"), val = bool(false)];
             bool context_layer_93_transpose_y_0 = const()[name = string("context_layer_93_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 256, 64]> context_layer_93_cast_fp16 = matmul(transpose_x = context_layer_93_transpose_x_0, transpose_y = context_layer_93_transpose_y_0, x = var_5383_cast_fp16, y = value_layer_cast_fp16)[name = string("context_layer_93_cast_fp16")];
-            tensor<int32, [4]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [4]>([-1, 16, 256, 64])];
-            tensor<fp16, [1, 16, 256, 64]> var_5388_cast_fp16 = reshape(shape = var_5387, x = context_layer_93_cast_fp16)[name = string("op_5388_cast_fp16")];
+            tensor<fp16, [16, 192, 64]> context_layer_93_cast_fp16 = matmul(transpose_x = context_layer_93_transpose_x_0, transpose_y = context_layer_93_transpose_y_0, x = var_5383_cast_fp16, y = value_layer_cast_fp16)[name = string("context_layer_93_cast_fp16")];
+            tensor<int32, [4]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [4]>([-1, 16, 192, 64])];
+            tensor<fp16, [1, 16, 192, 64]> var_5388_cast_fp16 = reshape(shape = var_5387, x = context_layer_93_cast_fp16)[name = string("op_5388_cast_fp16")];
             tensor<int32, [4]> var_5389 = const()[name = string("op_5389"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [3]>([1, 256, -1])];
-            tensor<fp16, [1, 256, 16, 64]> var_5390_cast_fp16 = transpose(perm = var_5389, x = var_5388_cast_fp16)[name = string("transpose_0")];
-            tensor<fp16, [1, 256, 1024]> input_565_cast_fp16 = reshape(shape = var_5394, x = var_5390_cast_fp16)[name = string("input_565_cast_fp16")];
+            tensor<int32, [3]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [3]>([1, 192, -1])];
+            tensor<fp16, [1, 192, 16, 64]> var_5390_cast_fp16 = transpose(perm = var_5389, x = var_5388_cast_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 192, 1024]> input_565_cast_fp16 = reshape(shape = var_5394, x = var_5390_cast_fp16)[name = string("input_565_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(899302720)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901399936)))];
-            tensor<fp16, [1, 256, 1024]> linear_189_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_attention_output_dense_weight_to_fp16, x = input_565_cast_fp16)[name = string("linear_189_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_569_cast_fp16 = add(x = linear_189_cast_fp16, y = input_557_cast_fp16)[name = string("input_569_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_189_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_attention_output_dense_weight_to_fp16, x = input_565_cast_fp16)[name = string("linear_189_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_569_cast_fp16 = add(x = linear_189_cast_fp16, y = input_557_cast_fp16)[name = string("input_569_cast_fp16")];
             tensor<int32, [1]> input_571_axes_0 = const()[name = string("input_571_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901402048)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901404160)))];
-            tensor<fp16, [1, 256, 1024]> input_571_cast_fp16 = layer_norm(axes = input_571_axes_0, beta = encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_571_cast_fp16 = layer_norm(axes = input_571_axes_0, beta = encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_23_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901406272)))];
             tensor<fp16, [4096]> encoder_encoder_layer_23_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(909794944)))];
-            tensor<fp16, [1, 256, 4096]> linear_190_cast_fp16 = linear(bias = encoder_encoder_layer_23_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_23_intermediate_dense_weight_to_fp16, x = input_571_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> linear_190_cast_fp16 = linear(bias = encoder_encoder_layer_23_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_23_intermediate_dense_weight_to_fp16, x = input_571_cast_fp16)[name = string("linear_190_cast_fp16")];
             string input_575_mode_0 = const()[name = string("input_575_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 256, 4096]> input_575_cast_fp16 = gelu(mode = input_575_mode_0, x = linear_190_cast_fp16)[name = string("input_575_cast_fp16")];
+            tensor<fp16, [1, 192, 4096]> input_575_cast_fp16 = gelu(mode = input_575_mode_0, x = linear_190_cast_fp16)[name = string("input_575_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_23_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(909803200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918191872)))];
-            tensor<fp16, [1, 256, 1024]> linear_191_cast_fp16 = linear(bias = encoder_encoder_layer_23_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_output_dense_weight_to_fp16, x = input_575_cast_fp16)[name = string("linear_191_cast_fp16")];
-            tensor<fp16, [1, 256, 1024]> input_579_cast_fp16 = add(x = linear_191_cast_fp16, y = input_571_cast_fp16)[name = string("input_579_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> linear_191_cast_fp16 = linear(bias = encoder_encoder_layer_23_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_output_dense_weight_to_fp16, x = input_575_cast_fp16)[name = string("linear_191_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> input_579_cast_fp16 = add(x = linear_191_cast_fp16, y = input_571_cast_fp16)[name = string("input_579_cast_fp16")];
             tensor<int32, [1]> hidden_axes_0 = const()[name = string("hidden_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918193984)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918196096)))];
-            tensor<fp16, [1, 256, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, beta = encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16, x = input_579_cast_fp16)[name = string("hidden_cast_fp16")];
+            tensor<fp16, [1, 192, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, beta = encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16, x = input_579_cast_fp16)[name = string("hidden_cast_fp16")];
             tensor<int32, [3]> var_5426 = const()[name = string("op_5426"), val = tensor<int32, [3]>([1, 16, 1])];
             tensor<int32, [1, 16, 1]> var_5427 = reshape(shape = var_5426, x = label_positions)[name = string("op_5427")];
-            tensor<bool, [1, 16, 256]> var_5428 = equal(x = var_5427, y = slots)[name = string("op_5428")];
+            tensor<bool, [1, 16, 192]> var_5428 = equal(x = var_5427, y = slots)[name = string("op_5428")];
             bool input_581_transpose_x_0 = const()[name = string("input_581_transpose_x_0"), val = bool(false)];
             bool input_581_transpose_y_0 = const()[name = string("input_581_transpose_y_0"), val = bool(false)];
             string onehot_to_fp16_dtype_0 = const()[name = string("onehot_to_fp16_dtype_0"), val = string("fp16")];
-            tensor<fp16, [1, 16, 256]> var_5428_to_fp16 = cast(dtype = onehot_to_fp16_dtype_0, x = var_5428)[name = string("cast_490")];
+            tensor<fp16, [1, 16, 192]> var_5428_to_fp16 = cast(dtype = onehot_to_fp16_dtype_0, x = var_5428)[name = string("cast_490")];
             tensor<fp16, [1, 16, 1024]> input_581_cast_fp16 = matmul(transpose_x = input_581_transpose_x_0, transpose_y = input_581_transpose_y_0, x = var_5428_to_fp16, y = hidden_cast_fp16)[name = string("input_581_cast_fp16")];
             tensor<fp16, [2048, 1024]> classifier_0_weight_to_fp16 = const()[name = string("classifier_0_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918198208)))];
             tensor<fp16, [2048]> classifier_0_bias_to_fp16 = const()[name = string("classifier_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922392576)))];
@@ -4433,2209 +4433,2209 @@ program(1.3)
             string var_5445_cast_fp16_to_fp32_dtype_0 = const()[name = string("op_5445_cast_fp16_to_fp32_dtype_0"), val = string("fp32")];
             tensor<fp32, [1, 16]> logits = cast(dtype = var_5445_cast_fp16_to_fp32_dtype_0, x = var_5445_cast_fp16)[name = string("cast_489")];
         } -> (logits);
-    func len512<ios18>(tensor<int32, [1, 512]> attention_mask, tensor<int32, [1, 512]> input_ids, tensor<int32, [1, 16]> label_positions) {
-            tensor<int32, [1, 1, 512]> slots = const()[name = string("slots"), val = tensor<int32, [1, 1, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(926596416)))];
+    func len224<ios18>(tensor<int32, [1, 224]> attention_mask, tensor<int32, [1, 224]> input_ids, tensor<int32, [1, 16]> label_positions) {
+            tensor<int32, [1, 1, 224]> slots = const()[name = string("slots"), val = tensor<int32, [1, 1, 224]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(924761152)))];
             int32 inputs_embeds_batch_dims_0 = const()[name = string("inputs_embeds_batch_dims_0"), val = int32(0)];
             bool inputs_embeds_validate_indices_0 = const()[name = string("inputs_embeds_validate_indices_0"), val = bool(false)];
             tensor<fp16, [128011, 1024]> encoder_embeddings_word_embeddings_weight_to_fp16 = const()[name = string("encoder_embeddings_word_embeddings_weight_to_fp16"), val = tensor<fp16, [128011, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(640)))];
             int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
-            tensor<bool, [1, 512]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
+            tensor<bool, [1, 224]> greater_equal_0 = greater_equal(x = input_ids, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
             int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(128011)];
-            tensor<int32, [1, 512]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
-            tensor<int32, [1, 512]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
+            tensor<int32, [1, 224]> add_0 = add(x = input_ids, y = slice_by_index_0)[name = string("add_0")];
+            tensor<int32, [1, 224]> select_0 = select(a = input_ids, b = add_0, cond = greater_equal_0)[name = string("select_0")];
             int32 inputs_embeds_cast_fp16_axis_0 = const()[name = string("inputs_embeds_cast_fp16_axis_0"), val = int32(0)];
-            tensor<fp16, [1, 512, 1024]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0, validate_indices = inputs_embeds_validate_indices_0, x = encoder_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> inputs_embeds_cast_fp16 = gather(axis = inputs_embeds_cast_fp16_axis_0, batch_dims = inputs_embeds_batch_dims_0, indices = select_0, validate_indices = inputs_embeds_validate_indices_0, x = encoder_embeddings_word_embeddings_weight_to_fp16)[name = string("inputs_embeds_cast_fp16")];
             tensor<int32, [1]> embeddings_1_axes_0 = const()[name = string("embeddings_1_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_embeddings_LayerNorm_weight_to_fp16 = const()[name = string("encoder_embeddings_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262167232)))];
             tensor<fp16, [1024]> encoder_embeddings_LayerNorm_bias_to_fp16 = const()[name = string("encoder_embeddings_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262169344)))];
             fp16 var_31_to_fp16 = const()[name = string("op_31_to_fp16"), val = fp16(0x1p-23)];
-            tensor<fp16, [1, 512, 1024]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = encoder_embeddings_LayerNorm_bias_to_fp16, epsilon = var_31_to_fp16, gamma = encoder_embeddings_LayerNorm_weight_to_fp16, x = inputs_embeds_cast_fp16)[name = string("embeddings_1_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> embeddings_1_cast_fp16 = layer_norm(axes = embeddings_1_axes_0, beta = encoder_embeddings_LayerNorm_bias_to_fp16, epsilon = var_31_to_fp16, gamma = encoder_embeddings_LayerNorm_weight_to_fp16, x = inputs_embeds_cast_fp16)[name = string("embeddings_1_cast_fp16")];
             tensor<int32, [1]> mask_3_axes_0 = const()[name = string("mask_3_axes_0"), val = tensor<int32, [1]>([2])];
-            tensor<int32, [1, 512, 1]> mask_3 = expand_dims(axes = mask_3_axes_0, x = attention_mask)[name = string("mask_3")];
+            tensor<int32, [1, 224, 1]> mask_3 = expand_dims(axes = mask_3_axes_0, x = attention_mask)[name = string("mask_3")];
             string mask_to_fp16_dtype_0 = const()[name = string("mask_to_fp16_dtype_0"), val = string("fp16")];
-            tensor<fp16, [1, 512, 1]> mask_3_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = mask_3)[name = string("cast_492")];
-            tensor<fp16, [1, 512, 1024]> input_1_cast_fp16 = mul(x = embeddings_1_cast_fp16, y = mask_3_to_fp16)[name = string("input_1_cast_fp16")];
-            tensor<int32, [3]> var_49 = const()[name = string("op_49"), val = tensor<int32, [3]>([1, 1, 512])];
-            tensor<int32, [1, 1, 512]> var_50 = reshape(shape = var_49, x = attention_mask)[name = string("op_50")];
-            tensor<int32, [3]> attention_mask_3_reps_0 = const()[name = string("attention_mask_3_reps_0"), val = tensor<int32, [3]>([1, 512, 1])];
-            tensor<int32, [1, 512, 512]> attention_mask_3 = tile(reps = attention_mask_3_reps_0, x = var_50)[name = string("attention_mask_3")];
+            tensor<fp16, [1, 224, 1]> mask_3_to_fp16 = cast(dtype = mask_to_fp16_dtype_0, x = mask_3)[name = string("cast_492")];
+            tensor<fp16, [1, 224, 1024]> input_1_cast_fp16 = mul(x = embeddings_1_cast_fp16, y = mask_3_to_fp16)[name = string("input_1_cast_fp16")];
+            tensor<int32, [3]> var_49 = const()[name = string("op_49"), val = tensor<int32, [3]>([1, 1, 224])];
+            tensor<int32, [1, 1, 224]> var_50 = reshape(shape = var_49, x = attention_mask)[name = string("op_50")];
+            tensor<int32, [3]> attention_mask_3_reps_0 = const()[name = string("attention_mask_3_reps_0"), val = tensor<int32, [3]>([1, 224, 1])];
+            tensor<int32, [1, 224, 224]> attention_mask_3 = tile(reps = attention_mask_3_reps_0, x = var_50)[name = string("attention_mask_3")];
             int32 var_56 = const()[name = string("op_56"), val = int32(-1)];
             tensor<int32, [1]> attention_mask_5_axes_0 = const()[name = string("attention_mask_5_axes_0"), val = tensor<int32, [1]>([1])];
-            tensor<int32, [1, 1, 512, 512]> attention_mask_5 = expand_dims(axes = attention_mask_5_axes_0, x = attention_mask_3)[name = string("attention_mask_5")];
+            tensor<int32, [1, 1, 224, 224]> attention_mask_5 = expand_dims(axes = attention_mask_5_axes_0, x = attention_mask_3)[name = string("attention_mask_5")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262171456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264268672)))];
-            tensor<fp16, [1, 512, 1024]> linear_0_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
-            tensor<int32, [4]> var_189 = const()[name = string("op_189"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_3_cast_fp16 = reshape(shape = var_189, x = linear_0_cast_fp16)[name = string("x_3_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_0_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_query_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
+            tensor<int32, [4]> var_189 = const()[name = string("op_189"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_3_cast_fp16 = reshape(shape = var_189, x = linear_0_cast_fp16)[name = string("x_3_cast_fp16")];
             tensor<int32, [4]> var_191 = const()[name = string("op_191"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_192_cast_fp16 = transpose(perm = var_191, x = x_3_cast_fp16)[name = string("transpose_143")];
-            tensor<fp16, [16, 512, 64]> query_layer_3_cast_fp16 = reshape(shape = var_196, x = var_192_cast_fp16)[name = string("query_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_196 = const()[name = string("op_196"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_192_cast_fp16 = transpose(perm = var_191, x = x_3_cast_fp16)[name = string("transpose_143")];
+            tensor<fp16, [16, 224, 64]> query_layer_3_cast_fp16 = reshape(shape = var_196, x = var_192_cast_fp16)[name = string("query_layer_3_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264270784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266368000)))];
-            tensor<fp16, [1, 512, 1024]> linear_1_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
-            tensor<int32, [4]> var_203 = const()[name = string("op_203"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_7_cast_fp16 = reshape(shape = var_203, x = linear_1_cast_fp16)[name = string("x_7_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_1_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_key_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
+            tensor<int32, [4]> var_203 = const()[name = string("op_203"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_7_cast_fp16 = reshape(shape = var_203, x = linear_1_cast_fp16)[name = string("x_7_cast_fp16")];
             tensor<int32, [4]> var_205 = const()[name = string("op_205"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_206_cast_fp16 = transpose(perm = var_205, x = x_7_cast_fp16)[name = string("transpose_142")];
-            tensor<fp16, [16, 512, 64]> key_layer_1_cast_fp16 = reshape(shape = var_210, x = var_206_cast_fp16)[name = string("key_layer_1_cast_fp16")];
+            tensor<int32, [3]> var_210 = const()[name = string("op_210"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_206_cast_fp16 = transpose(perm = var_205, x = x_7_cast_fp16)[name = string("transpose_142")];
+            tensor<fp16, [16, 224, 64]> key_layer_1_cast_fp16 = reshape(shape = var_210, x = var_206_cast_fp16)[name = string("key_layer_1_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266370112)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268467328)))];
-            tensor<fp16, [1, 512, 1024]> linear_2_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
-            tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_11_cast_fp16 = reshape(shape = var_217, x = linear_2_cast_fp16)[name = string("x_11_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_2_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_0_attention_self_value_proj_weight_to_fp16, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
+            tensor<int32, [4]> var_217 = const()[name = string("op_217"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_11_cast_fp16 = reshape(shape = var_217, x = linear_2_cast_fp16)[name = string("x_11_cast_fp16")];
             tensor<int32, [4]> var_219 = const()[name = string("op_219"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_224 = const()[name = string("op_224"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_220_cast_fp16 = transpose(perm = var_219, x = x_11_cast_fp16)[name = string("transpose_141")];
-            tensor<fp16, [16, 512, 64]> value_layer_1_cast_fp16 = reshape(shape = var_224, x = var_220_cast_fp16)[name = string("value_layer_1_cast_fp16")];
+            tensor<int32, [3]> var_224 = const()[name = string("op_224"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_220_cast_fp16 = transpose(perm = var_219, x = x_11_cast_fp16)[name = string("transpose_141")];
+            tensor<fp16, [16, 224, 64]> value_layer_1_cast_fp16 = reshape(shape = var_224, x = var_220_cast_fp16)[name = string("value_layer_1_cast_fp16")];
             tensor<int32, [3]> var_234_perm_0 = const()[name = string("op_234_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_236_y_0_to_fp16 = const()[name = string("_inversed_236_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_234_cast_fp16 = transpose(perm = var_234_perm_0, x = key_layer_1_cast_fp16)[name = string("transpose_140")];
-            tensor<fp16, [16, 64, 512]> _inversed_236_cast_fp16 = mul(x = var_234_cast_fp16, y = _inversed_236_y_0_to_fp16)[name = string("_inversed_236_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_234_cast_fp16 = transpose(perm = var_234_perm_0, x = key_layer_1_cast_fp16)[name = string("transpose_140")];
+            tensor<fp16, [16, 64, 224]> _inversed_236_cast_fp16 = mul(x = var_234_cast_fp16, y = _inversed_236_y_0_to_fp16)[name = string("_inversed_236_cast_fp16")];
             bool attention_scores_1_transpose_x_0 = const()[name = string("attention_scores_1_transpose_x_0"), val = bool(false)];
             bool attention_scores_1_transpose_y_0 = const()[name = string("attention_scores_1_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = query_layer_3_cast_fp16, y = _inversed_236_cast_fp16)[name = string("attention_scores_1_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_1_cast_fp16 = matmul(transpose_x = attention_scores_1_transpose_x_0, transpose_y = attention_scores_1_transpose_y_0, x = query_layer_3_cast_fp16, y = _inversed_236_cast_fp16)[name = string("attention_scores_1_cast_fp16")];
             bool c2p_att_1_transpose_x_0 = const()[name = string("c2p_att_1_transpose_x_0"), val = bool(false)];
             bool c2p_att_1_transpose_y_0 = const()[name = string("c2p_att_1_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_292_to_fp16 = const()[name = string("op_292_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268469440)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_1_cast_fp16 = matmul(transpose_x = c2p_att_1_transpose_x_0, transpose_y = c2p_att_1_transpose_y_0, x = query_layer_3_cast_fp16, y = var_292_to_fp16)[name = string("c2p_att_1_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_1_cast_fp16 = matmul(transpose_x = c2p_att_1_transpose_x_0, transpose_y = c2p_att_1_transpose_y_0, x = query_layer_3_cast_fp16, y = var_292_to_fp16)[name = string("c2p_att_1_cast_fp16")];
             bool c2p_att_3_validate_indices_0 = const()[name = string("c2p_att_3_validate_indices_0"), val = bool(false)];
-            tensor<uint16, [16, 512, 512]> var_302_to_uint16 = const()[name = string("op_302_to_uint16"), val = tensor<uint16, [16, 512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(926598528)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_3_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_3_validate_indices_0, x = c2p_att_1_cast_fp16)[name = string("c2p_att_3_cast_fp16_cast_int16")];
+            tensor<uint16, [16, 224, 224]> var_302_to_uint16 = const()[name = string("op_302_to_uint16"), val = tensor<uint16, [16, 224, 224]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(924762112)))];
+            tensor<fp16, [16, 224, 224]> c2p_att_3_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_3_validate_indices_0, x = c2p_att_1_cast_fp16)[name = string("c2p_att_3_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_305_y_0_to_fp16 = const()[name = string("_inversed_305_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_305_cast_fp16 = mul(x = c2p_att_3_cast_fp16_cast_int16, y = _inversed_305_y_0_to_fp16)[name = string("_inversed_305_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_305_cast_fp16 = mul(x = c2p_att_3_cast_fp16_cast_int16, y = _inversed_305_y_0_to_fp16)[name = string("_inversed_305_cast_fp16")];
             bool p2c_att_1_transpose_x_0 = const()[name = string("p2c_att_1_transpose_x_0"), val = bool(false)];
             bool p2c_att_1_transpose_y_0 = const()[name = string("p2c_att_1_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_320_to_fp16 = const()[name = string("op_320_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(270042432)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_1_cast_fp16 = matmul(transpose_x = p2c_att_1_transpose_x_0, transpose_y = p2c_att_1_transpose_y_0, x = key_layer_1_cast_fp16, y = var_320_to_fp16)[name = string("p2c_att_1_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_1_cast_fp16 = matmul(transpose_x = p2c_att_1_transpose_x_0, transpose_y = p2c_att_1_transpose_y_0, x = key_layer_1_cast_fp16, y = var_320_to_fp16)[name = string("p2c_att_1_cast_fp16")];
             bool var_328_validate_indices_0 = const()[name = string("op_328_validate_indices_0"), val = bool(false)];
-            tensor<uint16, [16, 512, 512]> var_327_to_uint16 = const()[name = string("op_327_to_uint16"), val = tensor<uint16, [16, 512, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(934987200)))];
-            tensor<fp16, [16, 512, 512]> var_328_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_328_validate_indices_0, x = p2c_att_1_cast_fp16)[name = string("op_328_cast_fp16_cast_int16")];
+            tensor<uint16, [16, 224, 224]> var_327_to_uint16 = const()[name = string("op_327_to_uint16"), val = tensor<uint16, [16, 224, 224]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(926367808)))];
+            tensor<fp16, [16, 224, 224]> var_328_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_328_validate_indices_0, x = p2c_att_1_cast_fp16)[name = string("op_328_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_3_perm_0 = const()[name = string("p2c_att_3_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_331_y_0_to_fp16 = const()[name = string("_inversed_331_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_3_cast_fp16 = transpose(perm = p2c_att_3_perm_0, x = var_328_cast_fp16_cast_int16)[name = string("transpose_139")];
-            tensor<fp16, [16, 512, 512]> _inversed_331_cast_fp16 = mul(x = p2c_att_3_cast_fp16, y = _inversed_331_y_0_to_fp16)[name = string("_inversed_331_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_1_cast_fp16 = add(x = _inversed_305_cast_fp16, y = _inversed_331_cast_fp16)[name = string("rel_att_1_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_3_cast_fp16 = add(x = attention_scores_1_cast_fp16, y = rel_att_1_cast_fp16)[name = string("attention_scores_3_cast_fp16")];
-            tensor<int32, [4]> var_336 = const()[name = string("op_336"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_5_cast_fp16 = reshape(shape = var_336, x = attention_scores_3_cast_fp16)[name = string("attention_scores_5_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_3_cast_fp16 = transpose(perm = p2c_att_3_perm_0, x = var_328_cast_fp16_cast_int16)[name = string("transpose_139")];
+            tensor<fp16, [16, 224, 224]> _inversed_331_cast_fp16 = mul(x = p2c_att_3_cast_fp16, y = _inversed_331_y_0_to_fp16)[name = string("_inversed_331_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_1_cast_fp16 = add(x = _inversed_305_cast_fp16, y = _inversed_331_cast_fp16)[name = string("rel_att_1_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_3_cast_fp16 = add(x = attention_scores_1_cast_fp16, y = rel_att_1_cast_fp16)[name = string("attention_scores_3_cast_fp16")];
+            tensor<int32, [4]> var_336 = const()[name = string("op_336"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_5_cast_fp16 = reshape(shape = var_336, x = attention_scores_3_cast_fp16)[name = string("attention_scores_5_cast_fp16")];
             string attention_mask_7_dtype_0 = const()[name = string("attention_mask_7_dtype_0"), val = string("bool")];
-            tensor<bool, [1, 1, 512, 512]> attention_mask_7 = cast(dtype = attention_mask_7_dtype_0, x = attention_mask_5)[name = string("cast_491")];
-            tensor<bool, [1, 1, 512, 512]> var_339 = logical_not(x = attention_mask_7)[name = string("op_339")];
+            tensor<bool, [1, 1, 224, 224]> attention_mask_7 = cast(dtype = attention_mask_7_dtype_0, x = attention_mask_5)[name = string("cast_491")];
+            tensor<bool, [1, 1, 224, 224]> var_339 = logical_not(x = attention_mask_7)[name = string("op_339")];
             fp16 var_64_to_fp16 = const()[name = string("op_64_to_fp16"), val = fp16(-inf)];
-            tensor<fp16, [1, 16, 512, 512]> input_9_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_5_cast_fp16, cond = var_339)[name = string("input_9_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_11_cast_fp16 = softmax(axis = var_56, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
-            tensor<int32, [3]> var_345 = const()[name = string("op_345"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_346_cast_fp16 = reshape(shape = var_345, x = input_11_cast_fp16)[name = string("op_346_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_9_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_5_cast_fp16, cond = var_339)[name = string("input_9_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_11_cast_fp16 = softmax(axis = var_56, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
+            tensor<int32, [3]> var_345 = const()[name = string("op_345"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_346_cast_fp16 = reshape(shape = var_345, x = input_11_cast_fp16)[name = string("op_346_cast_fp16")];
             bool context_layer_1_transpose_x_0 = const()[name = string("context_layer_1_transpose_x_0"), val = bool(false)];
             bool context_layer_1_transpose_y_0 = const()[name = string("context_layer_1_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = var_346_cast_fp16, y = value_layer_1_cast_fp16)[name = string("context_layer_1_cast_fp16")];
-            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_351_cast_fp16 = reshape(shape = var_350, x = context_layer_1_cast_fp16)[name = string("op_351_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_1_cast_fp16 = matmul(transpose_x = context_layer_1_transpose_x_0, transpose_y = context_layer_1_transpose_y_0, x = var_346_cast_fp16, y = value_layer_1_cast_fp16)[name = string("context_layer_1_cast_fp16")];
+            tensor<int32, [4]> var_350 = const()[name = string("op_350"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_351_cast_fp16 = reshape(shape = var_350, x = context_layer_1_cast_fp16)[name = string("op_351_cast_fp16")];
             tensor<int32, [4]> var_352 = const()[name = string("op_352"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_357 = const()[name = string("op_357"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_353_cast_fp16 = transpose(perm = var_352, x = var_351_cast_fp16)[name = string("transpose_138")];
-            tensor<fp16, [1, 512, 1024]> input_13_cast_fp16 = reshape(shape = var_357, x = var_353_cast_fp16)[name = string("input_13_cast_fp16")];
+            tensor<int32, [3]> var_357 = const()[name = string("op_357"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_353_cast_fp16 = transpose(perm = var_352, x = var_351_cast_fp16)[name = string("transpose_138")];
+            tensor<fp16, [1, 224, 1024]> input_13_cast_fp16 = reshape(shape = var_357, x = var_353_cast_fp16)[name = string("input_13_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_0_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(271615424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273712640)))];
-            tensor<fp16, [1, 512, 1024]> linear_5_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_5_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_17_cast_fp16 = add(x = linear_5_cast_fp16, y = input_1_cast_fp16)[name = string("input_17_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_5_cast_fp16 = linear(bias = encoder_encoder_layer_0_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_attention_output_dense_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_5_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_17_cast_fp16 = add(x = linear_5_cast_fp16, y = input_1_cast_fp16)[name = string("input_17_cast_fp16")];
             tensor<int32, [1]> input_19_axes_0 = const()[name = string("input_19_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273714752)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273716864)))];
             fp16 var_67_to_fp16 = const()[name = string("op_67_to_fp16"), val = fp16(0x1p-23)];
-            tensor<fp16, [1, 512, 1024]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_19_cast_fp16 = layer_norm(axes = input_19_axes_0, beta = encoder_encoder_layer_0_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_attention_output_LayerNorm_weight_to_fp16, x = input_17_cast_fp16)[name = string("input_19_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_0_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(273718976)))];
             tensor<fp16, [4096]> encoder_encoder_layer_0_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282107648)))];
-            tensor<fp16, [1, 512, 4096]> linear_6_cast_fp16 = linear(bias = encoder_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_6_cast_fp16 = linear(bias = encoder_encoder_layer_0_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_0_intermediate_dense_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_6_cast_fp16")];
             string input_23_mode_0 = const()[name = string("input_23_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = string("input_23_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_23_cast_fp16 = gelu(mode = input_23_mode_0, x = linear_6_cast_fp16)[name = string("input_23_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_0_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(282115904)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290504576)))];
-            tensor<fp16, [1, 512, 1024]> linear_7_cast_fp16 = linear(bias = encoder_encoder_layer_0_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_output_dense_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_27_cast_fp16 = add(x = linear_7_cast_fp16, y = input_19_cast_fp16)[name = string("input_27_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_7_cast_fp16 = linear(bias = encoder_encoder_layer_0_output_dense_bias_to_fp16, weight = encoder_encoder_layer_0_output_dense_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_7_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_27_cast_fp16 = add(x = linear_7_cast_fp16, y = input_19_cast_fp16)[name = string("input_27_cast_fp16")];
             tensor<int32, [1]> input_29_axes_0 = const()[name = string("input_29_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290506688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290508800)))];
-            tensor<fp16, [1, 512, 1024]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_29_cast_fp16 = layer_norm(axes = input_29_axes_0, beta = encoder_encoder_layer_0_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_0_output_LayerNorm_weight_to_fp16, x = input_27_cast_fp16)[name = string("input_29_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(290510912)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292608128)))];
-            tensor<fp16, [1, 512, 1024]> linear_8_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_8_cast_fp16")];
-            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_23_cast_fp16 = reshape(shape = var_408, x = linear_8_cast_fp16)[name = string("x_23_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_8_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_query_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_8_cast_fp16")];
+            tensor<int32, [4]> var_408 = const()[name = string("op_408"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_23_cast_fp16 = reshape(shape = var_408, x = linear_8_cast_fp16)[name = string("x_23_cast_fp16")];
             tensor<int32, [4]> var_410 = const()[name = string("op_410"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_411_cast_fp16 = transpose(perm = var_410, x = x_23_cast_fp16)[name = string("transpose_137")];
-            tensor<fp16, [16, 512, 64]> query_layer_9_cast_fp16 = reshape(shape = var_415, x = var_411_cast_fp16)[name = string("query_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_415 = const()[name = string("op_415"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_411_cast_fp16 = transpose(perm = var_410, x = x_23_cast_fp16)[name = string("transpose_137")];
+            tensor<fp16, [16, 224, 64]> query_layer_9_cast_fp16 = reshape(shape = var_415, x = var_411_cast_fp16)[name = string("query_layer_9_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(292610240)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294707456)))];
-            tensor<fp16, [1, 512, 1024]> linear_9_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_9_cast_fp16")];
-            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_27_cast_fp16 = reshape(shape = var_422, x = linear_9_cast_fp16)[name = string("x_27_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_9_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_key_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_9_cast_fp16")];
+            tensor<int32, [4]> var_422 = const()[name = string("op_422"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_27_cast_fp16 = reshape(shape = var_422, x = linear_9_cast_fp16)[name = string("x_27_cast_fp16")];
             tensor<int32, [4]> var_424 = const()[name = string("op_424"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_425_cast_fp16 = transpose(perm = var_424, x = x_27_cast_fp16)[name = string("transpose_136")];
-            tensor<fp16, [16, 512, 64]> key_layer_3_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = string("key_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_429 = const()[name = string("op_429"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_425_cast_fp16 = transpose(perm = var_424, x = x_27_cast_fp16)[name = string("transpose_136")];
+            tensor<fp16, [16, 224, 64]> key_layer_3_cast_fp16 = reshape(shape = var_429, x = var_425_cast_fp16)[name = string("key_layer_3_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(294709568)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296806784)))];
-            tensor<fp16, [1, 512, 1024]> linear_10_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_10_cast_fp16")];
-            tensor<int32, [4]> var_436 = const()[name = string("op_436"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_31_cast_fp16 = reshape(shape = var_436, x = linear_10_cast_fp16)[name = string("x_31_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_10_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_1_attention_self_value_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_10_cast_fp16")];
+            tensor<int32, [4]> var_436 = const()[name = string("op_436"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_31_cast_fp16 = reshape(shape = var_436, x = linear_10_cast_fp16)[name = string("x_31_cast_fp16")];
             tensor<int32, [4]> var_438 = const()[name = string("op_438"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_443 = const()[name = string("op_443"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_439_cast_fp16 = transpose(perm = var_438, x = x_31_cast_fp16)[name = string("transpose_135")];
-            tensor<fp16, [16, 512, 64]> value_layer_3_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = string("value_layer_3_cast_fp16")];
+            tensor<int32, [3]> var_443 = const()[name = string("op_443"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_439_cast_fp16 = transpose(perm = var_438, x = x_31_cast_fp16)[name = string("transpose_135")];
+            tensor<fp16, [16, 224, 64]> value_layer_3_cast_fp16 = reshape(shape = var_443, x = var_439_cast_fp16)[name = string("value_layer_3_cast_fp16")];
             tensor<int32, [3]> var_453_perm_0 = const()[name = string("op_453_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_455_y_0_to_fp16 = const()[name = string("_inversed_455_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_453_cast_fp16 = transpose(perm = var_453_perm_0, x = key_layer_3_cast_fp16)[name = string("transpose_134")];
-            tensor<fp16, [16, 64, 512]> _inversed_455_cast_fp16 = mul(x = var_453_cast_fp16, y = _inversed_455_y_0_to_fp16)[name = string("_inversed_455_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_453_cast_fp16 = transpose(perm = var_453_perm_0, x = key_layer_3_cast_fp16)[name = string("transpose_134")];
+            tensor<fp16, [16, 64, 224]> _inversed_455_cast_fp16 = mul(x = var_453_cast_fp16, y = _inversed_455_y_0_to_fp16)[name = string("_inversed_455_cast_fp16")];
             bool attention_scores_7_transpose_x_0 = const()[name = string("attention_scores_7_transpose_x_0"), val = bool(false)];
             bool attention_scores_7_transpose_y_0 = const()[name = string("attention_scores_7_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = query_layer_9_cast_fp16, y = _inversed_455_cast_fp16)[name = string("attention_scores_7_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_7_cast_fp16 = matmul(transpose_x = attention_scores_7_transpose_x_0, transpose_y = attention_scores_7_transpose_y_0, x = query_layer_9_cast_fp16, y = _inversed_455_cast_fp16)[name = string("attention_scores_7_cast_fp16")];
             bool c2p_att_5_transpose_x_0 = const()[name = string("c2p_att_5_transpose_x_0"), val = bool(false)];
             bool c2p_att_5_transpose_y_0 = const()[name = string("c2p_att_5_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_511_to_fp16 = const()[name = string("op_511_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(296808896)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_5_cast_fp16 = matmul(transpose_x = c2p_att_5_transpose_x_0, transpose_y = c2p_att_5_transpose_y_0, x = query_layer_9_cast_fp16, y = var_511_to_fp16)[name = string("c2p_att_5_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_5_cast_fp16 = matmul(transpose_x = c2p_att_5_transpose_x_0, transpose_y = c2p_att_5_transpose_y_0, x = query_layer_9_cast_fp16, y = var_511_to_fp16)[name = string("c2p_att_5_cast_fp16")];
             bool c2p_att_7_validate_indices_0 = const()[name = string("c2p_att_7_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_7_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_7_validate_indices_0, x = c2p_att_5_cast_fp16)[name = string("c2p_att_7_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_7_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_7_validate_indices_0, x = c2p_att_5_cast_fp16)[name = string("c2p_att_7_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_524_y_0_to_fp16 = const()[name = string("_inversed_524_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_524_cast_fp16 = mul(x = c2p_att_7_cast_fp16_cast_int16, y = _inversed_524_y_0_to_fp16)[name = string("_inversed_524_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_524_cast_fp16 = mul(x = c2p_att_7_cast_fp16_cast_int16, y = _inversed_524_y_0_to_fp16)[name = string("_inversed_524_cast_fp16")];
             bool p2c_att_5_transpose_x_0 = const()[name = string("p2c_att_5_transpose_x_0"), val = bool(false)];
             bool p2c_att_5_transpose_y_0 = const()[name = string("p2c_att_5_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_539_to_fp16 = const()[name = string("op_539_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297857536)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_5_cast_fp16 = matmul(transpose_x = p2c_att_5_transpose_x_0, transpose_y = p2c_att_5_transpose_y_0, x = key_layer_3_cast_fp16, y = var_539_to_fp16)[name = string("p2c_att_5_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_5_cast_fp16 = matmul(transpose_x = p2c_att_5_transpose_x_0, transpose_y = p2c_att_5_transpose_y_0, x = key_layer_3_cast_fp16, y = var_539_to_fp16)[name = string("p2c_att_5_cast_fp16")];
             bool var_547_validate_indices_0 = const()[name = string("op_547_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_547_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_547_validate_indices_0, x = p2c_att_5_cast_fp16)[name = string("op_547_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_547_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_547_validate_indices_0, x = p2c_att_5_cast_fp16)[name = string("op_547_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_7_perm_0 = const()[name = string("p2c_att_7_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_550_y_0_to_fp16 = const()[name = string("_inversed_550_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_7_cast_fp16 = transpose(perm = p2c_att_7_perm_0, x = var_547_cast_fp16_cast_int16)[name = string("transpose_133")];
-            tensor<fp16, [16, 512, 512]> _inversed_550_cast_fp16 = mul(x = p2c_att_7_cast_fp16, y = _inversed_550_y_0_to_fp16)[name = string("_inversed_550_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_3_cast_fp16 = add(x = _inversed_524_cast_fp16, y = _inversed_550_cast_fp16)[name = string("rel_att_3_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_9_cast_fp16 = add(x = attention_scores_7_cast_fp16, y = rel_att_3_cast_fp16)[name = string("attention_scores_9_cast_fp16")];
-            tensor<int32, [4]> var_555 = const()[name = string("op_555"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_11_cast_fp16 = reshape(shape = var_555, x = attention_scores_9_cast_fp16)[name = string("attention_scores_11_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_33_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_11_cast_fp16, cond = var_339)[name = string("input_33_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_35_cast_fp16 = softmax(axis = var_56, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
-            tensor<int32, [3]> var_564 = const()[name = string("op_564"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_565_cast_fp16 = reshape(shape = var_564, x = input_35_cast_fp16)[name = string("op_565_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_7_cast_fp16 = transpose(perm = p2c_att_7_perm_0, x = var_547_cast_fp16_cast_int16)[name = string("transpose_133")];
+            tensor<fp16, [16, 224, 224]> _inversed_550_cast_fp16 = mul(x = p2c_att_7_cast_fp16, y = _inversed_550_y_0_to_fp16)[name = string("_inversed_550_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_3_cast_fp16 = add(x = _inversed_524_cast_fp16, y = _inversed_550_cast_fp16)[name = string("rel_att_3_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_9_cast_fp16 = add(x = attention_scores_7_cast_fp16, y = rel_att_3_cast_fp16)[name = string("attention_scores_9_cast_fp16")];
+            tensor<int32, [4]> var_555 = const()[name = string("op_555"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_11_cast_fp16 = reshape(shape = var_555, x = attention_scores_9_cast_fp16)[name = string("attention_scores_11_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_33_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_11_cast_fp16, cond = var_339)[name = string("input_33_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_35_cast_fp16 = softmax(axis = var_56, x = input_33_cast_fp16)[name = string("input_35_cast_fp16")];
+            tensor<int32, [3]> var_564 = const()[name = string("op_564"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_565_cast_fp16 = reshape(shape = var_564, x = input_35_cast_fp16)[name = string("op_565_cast_fp16")];
             bool context_layer_5_transpose_x_0 = const()[name = string("context_layer_5_transpose_x_0"), val = bool(false)];
             bool context_layer_5_transpose_y_0 = const()[name = string("context_layer_5_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = var_565_cast_fp16, y = value_layer_3_cast_fp16)[name = string("context_layer_5_cast_fp16")];
-            tensor<int32, [4]> var_569 = const()[name = string("op_569"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_570_cast_fp16 = reshape(shape = var_569, x = context_layer_5_cast_fp16)[name = string("op_570_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_5_cast_fp16 = matmul(transpose_x = context_layer_5_transpose_x_0, transpose_y = context_layer_5_transpose_y_0, x = var_565_cast_fp16, y = value_layer_3_cast_fp16)[name = string("context_layer_5_cast_fp16")];
+            tensor<int32, [4]> var_569 = const()[name = string("op_569"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_570_cast_fp16 = reshape(shape = var_569, x = context_layer_5_cast_fp16)[name = string("op_570_cast_fp16")];
             tensor<int32, [4]> var_571 = const()[name = string("op_571"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_576 = const()[name = string("op_576"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_572_cast_fp16 = transpose(perm = var_571, x = var_570_cast_fp16)[name = string("transpose_132")];
-            tensor<fp16, [1, 512, 1024]> input_37_cast_fp16 = reshape(shape = var_576, x = var_572_cast_fp16)[name = string("input_37_cast_fp16")];
+            tensor<int32, [3]> var_576 = const()[name = string("op_576"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_572_cast_fp16 = transpose(perm = var_571, x = var_570_cast_fp16)[name = string("transpose_132")];
+            tensor<fp16, [1, 224, 1024]> input_37_cast_fp16 = reshape(shape = var_576, x = var_572_cast_fp16)[name = string("input_37_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_1_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(298906176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301003392)))];
-            tensor<fp16, [1, 512, 1024]> linear_13_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_13_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_41_cast_fp16 = add(x = linear_13_cast_fp16, y = input_29_cast_fp16)[name = string("input_41_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_13_cast_fp16 = linear(bias = encoder_encoder_layer_1_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_attention_output_dense_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_13_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_41_cast_fp16 = add(x = linear_13_cast_fp16, y = input_29_cast_fp16)[name = string("input_41_cast_fp16")];
             tensor<int32, [1]> input_43_axes_0 = const()[name = string("input_43_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301005504)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301007616)))];
-            tensor<fp16, [1, 512, 1024]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_43_cast_fp16 = layer_norm(axes = input_43_axes_0, beta = encoder_encoder_layer_1_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_attention_output_LayerNorm_weight_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_1_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(301009728)))];
             tensor<fp16, [4096]> encoder_encoder_layer_1_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309398400)))];
-            tensor<fp16, [1, 512, 4096]> linear_14_cast_fp16 = linear(bias = encoder_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_14_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_14_cast_fp16 = linear(bias = encoder_encoder_layer_1_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_1_intermediate_dense_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_14_cast_fp16")];
             string input_47_mode_0 = const()[name = string("input_47_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = string("input_47_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_47_cast_fp16 = gelu(mode = input_47_mode_0, x = linear_14_cast_fp16)[name = string("input_47_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_1_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(309406656)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317795328)))];
-            tensor<fp16, [1, 512, 1024]> linear_15_cast_fp16 = linear(bias = encoder_encoder_layer_1_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_output_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_15_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_51_cast_fp16 = add(x = linear_15_cast_fp16, y = input_43_cast_fp16)[name = string("input_51_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_15_cast_fp16 = linear(bias = encoder_encoder_layer_1_output_dense_bias_to_fp16, weight = encoder_encoder_layer_1_output_dense_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_15_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_51_cast_fp16 = add(x = linear_15_cast_fp16, y = input_43_cast_fp16)[name = string("input_51_cast_fp16")];
             tensor<int32, [1]> input_53_axes_0 = const()[name = string("input_53_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317797440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317799552)))];
-            tensor<fp16, [1, 512, 1024]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_53_cast_fp16 = layer_norm(axes = input_53_axes_0, beta = encoder_encoder_layer_1_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_1_output_LayerNorm_weight_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317801664)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319898880)))];
-            tensor<fp16, [1, 512, 1024]> linear_16_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_16_cast_fp16")];
-            tensor<int32, [4]> var_627 = const()[name = string("op_627"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_43_cast_fp16 = reshape(shape = var_627, x = linear_16_cast_fp16)[name = string("x_43_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_16_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_query_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_16_cast_fp16")];
+            tensor<int32, [4]> var_627 = const()[name = string("op_627"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_43_cast_fp16 = reshape(shape = var_627, x = linear_16_cast_fp16)[name = string("x_43_cast_fp16")];
             tensor<int32, [4]> var_629 = const()[name = string("op_629"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_634 = const()[name = string("op_634"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_630_cast_fp16 = transpose(perm = var_629, x = x_43_cast_fp16)[name = string("transpose_131")];
-            tensor<fp16, [16, 512, 64]> query_layer_15_cast_fp16 = reshape(shape = var_634, x = var_630_cast_fp16)[name = string("query_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_634 = const()[name = string("op_634"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_630_cast_fp16 = transpose(perm = var_629, x = x_43_cast_fp16)[name = string("transpose_131")];
+            tensor<fp16, [16, 224, 64]> query_layer_15_cast_fp16 = reshape(shape = var_634, x = var_630_cast_fp16)[name = string("query_layer_15_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(319900992)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321998208)))];
-            tensor<fp16, [1, 512, 1024]> linear_17_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_17_cast_fp16")];
-            tensor<int32, [4]> var_641 = const()[name = string("op_641"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_47_cast_fp16 = reshape(shape = var_641, x = linear_17_cast_fp16)[name = string("x_47_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_17_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_key_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_17_cast_fp16")];
+            tensor<int32, [4]> var_641 = const()[name = string("op_641"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_47_cast_fp16 = reshape(shape = var_641, x = linear_17_cast_fp16)[name = string("x_47_cast_fp16")];
             tensor<int32, [4]> var_643 = const()[name = string("op_643"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_648 = const()[name = string("op_648"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_644_cast_fp16 = transpose(perm = var_643, x = x_47_cast_fp16)[name = string("transpose_130")];
-            tensor<fp16, [16, 512, 64]> key_layer_5_cast_fp16 = reshape(shape = var_648, x = var_644_cast_fp16)[name = string("key_layer_5_cast_fp16")];
+            tensor<int32, [3]> var_648 = const()[name = string("op_648"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_644_cast_fp16 = transpose(perm = var_643, x = x_47_cast_fp16)[name = string("transpose_130")];
+            tensor<fp16, [16, 224, 64]> key_layer_5_cast_fp16 = reshape(shape = var_648, x = var_644_cast_fp16)[name = string("key_layer_5_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(322000320)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324097536)))];
-            tensor<fp16, [1, 512, 1024]> linear_18_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_18_cast_fp16")];
-            tensor<int32, [4]> var_655 = const()[name = string("op_655"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_51_cast_fp16 = reshape(shape = var_655, x = linear_18_cast_fp16)[name = string("x_51_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_18_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_2_attention_self_value_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_18_cast_fp16")];
+            tensor<int32, [4]> var_655 = const()[name = string("op_655"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_51_cast_fp16 = reshape(shape = var_655, x = linear_18_cast_fp16)[name = string("x_51_cast_fp16")];
             tensor<int32, [4]> var_657 = const()[name = string("op_657"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_662 = const()[name = string("op_662"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_658_cast_fp16 = transpose(perm = var_657, x = x_51_cast_fp16)[name = string("transpose_129")];
-            tensor<fp16, [16, 512, 64]> value_layer_5_cast_fp16 = reshape(shape = var_662, x = var_658_cast_fp16)[name = string("value_layer_5_cast_fp16")];
+            tensor<int32, [3]> var_662 = const()[name = string("op_662"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_658_cast_fp16 = transpose(perm = var_657, x = x_51_cast_fp16)[name = string("transpose_129")];
+            tensor<fp16, [16, 224, 64]> value_layer_5_cast_fp16 = reshape(shape = var_662, x = var_658_cast_fp16)[name = string("value_layer_5_cast_fp16")];
             tensor<int32, [3]> var_672_perm_0 = const()[name = string("op_672_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_674_y_0_to_fp16 = const()[name = string("_inversed_674_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_672_cast_fp16 = transpose(perm = var_672_perm_0, x = key_layer_5_cast_fp16)[name = string("transpose_128")];
-            tensor<fp16, [16, 64, 512]> _inversed_674_cast_fp16 = mul(x = var_672_cast_fp16, y = _inversed_674_y_0_to_fp16)[name = string("_inversed_674_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_672_cast_fp16 = transpose(perm = var_672_perm_0, x = key_layer_5_cast_fp16)[name = string("transpose_128")];
+            tensor<fp16, [16, 64, 224]> _inversed_674_cast_fp16 = mul(x = var_672_cast_fp16, y = _inversed_674_y_0_to_fp16)[name = string("_inversed_674_cast_fp16")];
             bool attention_scores_13_transpose_x_0 = const()[name = string("attention_scores_13_transpose_x_0"), val = bool(false)];
             bool attention_scores_13_transpose_y_0 = const()[name = string("attention_scores_13_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = query_layer_15_cast_fp16, y = _inversed_674_cast_fp16)[name = string("attention_scores_13_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_13_cast_fp16 = matmul(transpose_x = attention_scores_13_transpose_x_0, transpose_y = attention_scores_13_transpose_y_0, x = query_layer_15_cast_fp16, y = _inversed_674_cast_fp16)[name = string("attention_scores_13_cast_fp16")];
             bool c2p_att_9_transpose_x_0 = const()[name = string("c2p_att_9_transpose_x_0"), val = bool(false)];
             bool c2p_att_9_transpose_y_0 = const()[name = string("c2p_att_9_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_730_to_fp16 = const()[name = string("op_730_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(324099648)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_9_cast_fp16 = matmul(transpose_x = c2p_att_9_transpose_x_0, transpose_y = c2p_att_9_transpose_y_0, x = query_layer_15_cast_fp16, y = var_730_to_fp16)[name = string("c2p_att_9_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_9_cast_fp16 = matmul(transpose_x = c2p_att_9_transpose_x_0, transpose_y = c2p_att_9_transpose_y_0, x = query_layer_15_cast_fp16, y = var_730_to_fp16)[name = string("c2p_att_9_cast_fp16")];
             bool c2p_att_11_validate_indices_0 = const()[name = string("c2p_att_11_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_11_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_11_validate_indices_0, x = c2p_att_9_cast_fp16)[name = string("c2p_att_11_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_11_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_11_validate_indices_0, x = c2p_att_9_cast_fp16)[name = string("c2p_att_11_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_743_y_0_to_fp16 = const()[name = string("_inversed_743_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_743_cast_fp16 = mul(x = c2p_att_11_cast_fp16_cast_int16, y = _inversed_743_y_0_to_fp16)[name = string("_inversed_743_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_743_cast_fp16 = mul(x = c2p_att_11_cast_fp16_cast_int16, y = _inversed_743_y_0_to_fp16)[name = string("_inversed_743_cast_fp16")];
             bool p2c_att_9_transpose_x_0 = const()[name = string("p2c_att_9_transpose_x_0"), val = bool(false)];
             bool p2c_att_9_transpose_y_0 = const()[name = string("p2c_att_9_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_758_to_fp16 = const()[name = string("op_758_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(325148288)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_9_cast_fp16 = matmul(transpose_x = p2c_att_9_transpose_x_0, transpose_y = p2c_att_9_transpose_y_0, x = key_layer_5_cast_fp16, y = var_758_to_fp16)[name = string("p2c_att_9_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_9_cast_fp16 = matmul(transpose_x = p2c_att_9_transpose_x_0, transpose_y = p2c_att_9_transpose_y_0, x = key_layer_5_cast_fp16, y = var_758_to_fp16)[name = string("p2c_att_9_cast_fp16")];
             bool var_766_validate_indices_0 = const()[name = string("op_766_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_766_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_766_validate_indices_0, x = p2c_att_9_cast_fp16)[name = string("op_766_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_766_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_766_validate_indices_0, x = p2c_att_9_cast_fp16)[name = string("op_766_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_11_perm_0 = const()[name = string("p2c_att_11_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_769_y_0_to_fp16 = const()[name = string("_inversed_769_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_11_cast_fp16 = transpose(perm = p2c_att_11_perm_0, x = var_766_cast_fp16_cast_int16)[name = string("transpose_127")];
-            tensor<fp16, [16, 512, 512]> _inversed_769_cast_fp16 = mul(x = p2c_att_11_cast_fp16, y = _inversed_769_y_0_to_fp16)[name = string("_inversed_769_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_5_cast_fp16 = add(x = _inversed_743_cast_fp16, y = _inversed_769_cast_fp16)[name = string("rel_att_5_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = rel_att_5_cast_fp16)[name = string("attention_scores_15_cast_fp16")];
-            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_17_cast_fp16 = reshape(shape = var_774, x = attention_scores_15_cast_fp16)[name = string("attention_scores_17_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_57_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_17_cast_fp16, cond = var_339)[name = string("input_57_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_59_cast_fp16 = softmax(axis = var_56, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
-            tensor<int32, [3]> var_783 = const()[name = string("op_783"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_784_cast_fp16 = reshape(shape = var_783, x = input_59_cast_fp16)[name = string("op_784_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_11_cast_fp16 = transpose(perm = p2c_att_11_perm_0, x = var_766_cast_fp16_cast_int16)[name = string("transpose_127")];
+            tensor<fp16, [16, 224, 224]> _inversed_769_cast_fp16 = mul(x = p2c_att_11_cast_fp16, y = _inversed_769_y_0_to_fp16)[name = string("_inversed_769_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_5_cast_fp16 = add(x = _inversed_743_cast_fp16, y = _inversed_769_cast_fp16)[name = string("rel_att_5_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_15_cast_fp16 = add(x = attention_scores_13_cast_fp16, y = rel_att_5_cast_fp16)[name = string("attention_scores_15_cast_fp16")];
+            tensor<int32, [4]> var_774 = const()[name = string("op_774"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_17_cast_fp16 = reshape(shape = var_774, x = attention_scores_15_cast_fp16)[name = string("attention_scores_17_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_57_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_17_cast_fp16, cond = var_339)[name = string("input_57_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_59_cast_fp16 = softmax(axis = var_56, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
+            tensor<int32, [3]> var_783 = const()[name = string("op_783"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_784_cast_fp16 = reshape(shape = var_783, x = input_59_cast_fp16)[name = string("op_784_cast_fp16")];
             bool context_layer_9_transpose_x_0 = const()[name = string("context_layer_9_transpose_x_0"), val = bool(false)];
             bool context_layer_9_transpose_y_0 = const()[name = string("context_layer_9_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = var_784_cast_fp16, y = value_layer_5_cast_fp16)[name = string("context_layer_9_cast_fp16")];
-            tensor<int32, [4]> var_788 = const()[name = string("op_788"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_789_cast_fp16 = reshape(shape = var_788, x = context_layer_9_cast_fp16)[name = string("op_789_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_9_cast_fp16 = matmul(transpose_x = context_layer_9_transpose_x_0, transpose_y = context_layer_9_transpose_y_0, x = var_784_cast_fp16, y = value_layer_5_cast_fp16)[name = string("context_layer_9_cast_fp16")];
+            tensor<int32, [4]> var_788 = const()[name = string("op_788"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_789_cast_fp16 = reshape(shape = var_788, x = context_layer_9_cast_fp16)[name = string("op_789_cast_fp16")];
             tensor<int32, [4]> var_790 = const()[name = string("op_790"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_795 = const()[name = string("op_795"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_791_cast_fp16 = transpose(perm = var_790, x = var_789_cast_fp16)[name = string("transpose_126")];
-            tensor<fp16, [1, 512, 1024]> input_61_cast_fp16 = reshape(shape = var_795, x = var_791_cast_fp16)[name = string("input_61_cast_fp16")];
+            tensor<int32, [3]> var_795 = const()[name = string("op_795"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_791_cast_fp16 = transpose(perm = var_790, x = var_789_cast_fp16)[name = string("transpose_126")];
+            tensor<fp16, [1, 224, 1024]> input_61_cast_fp16 = reshape(shape = var_795, x = var_791_cast_fp16)[name = string("input_61_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_2_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(326196928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328294144)))];
-            tensor<fp16, [1, 512, 1024]> linear_21_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_21_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_53_cast_fp16)[name = string("input_65_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_21_cast_fp16 = linear(bias = encoder_encoder_layer_2_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_attention_output_dense_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_21_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_65_cast_fp16 = add(x = linear_21_cast_fp16, y = input_53_cast_fp16)[name = string("input_65_cast_fp16")];
             tensor<int32, [1]> input_67_axes_0 = const()[name = string("input_67_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328296256)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328298368)))];
-            tensor<fp16, [1, 512, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_67_cast_fp16 = layer_norm(axes = input_67_axes_0, beta = encoder_encoder_layer_2_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_attention_output_LayerNorm_weight_to_fp16, x = input_65_cast_fp16)[name = string("input_67_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_2_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(328300480)))];
             tensor<fp16, [4096]> encoder_encoder_layer_2_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336689152)))];
-            tensor<fp16, [1, 512, 4096]> linear_22_cast_fp16 = linear(bias = encoder_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_22_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_22_cast_fp16 = linear(bias = encoder_encoder_layer_2_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_2_intermediate_dense_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_22_cast_fp16")];
             string input_71_mode_0 = const()[name = string("input_71_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = string("input_71_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_71_cast_fp16 = gelu(mode = input_71_mode_0, x = linear_22_cast_fp16)[name = string("input_71_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_2_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336697408)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345086080)))];
-            tensor<fp16, [1, 512, 1024]> linear_23_cast_fp16 = linear(bias = encoder_encoder_layer_2_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_output_dense_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_23_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_67_cast_fp16)[name = string("input_75_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_23_cast_fp16 = linear(bias = encoder_encoder_layer_2_output_dense_bias_to_fp16, weight = encoder_encoder_layer_2_output_dense_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_23_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_75_cast_fp16 = add(x = linear_23_cast_fp16, y = input_67_cast_fp16)[name = string("input_75_cast_fp16")];
             tensor<int32, [1]> input_77_axes_0 = const()[name = string("input_77_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345088192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345090304)))];
-            tensor<fp16, [1, 512, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_77_cast_fp16 = layer_norm(axes = input_77_axes_0, beta = encoder_encoder_layer_2_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_2_output_LayerNorm_weight_to_fp16, x = input_75_cast_fp16)[name = string("input_77_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345092416)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347189632)))];
-            tensor<fp16, [1, 512, 1024]> linear_24_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_24_cast_fp16")];
-            tensor<int32, [4]> var_846 = const()[name = string("op_846"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_63_cast_fp16 = reshape(shape = var_846, x = linear_24_cast_fp16)[name = string("x_63_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_24_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_query_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_24_cast_fp16")];
+            tensor<int32, [4]> var_846 = const()[name = string("op_846"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_63_cast_fp16 = reshape(shape = var_846, x = linear_24_cast_fp16)[name = string("x_63_cast_fp16")];
             tensor<int32, [4]> var_848 = const()[name = string("op_848"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_853 = const()[name = string("op_853"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_849_cast_fp16 = transpose(perm = var_848, x = x_63_cast_fp16)[name = string("transpose_125")];
-            tensor<fp16, [16, 512, 64]> query_layer_21_cast_fp16 = reshape(shape = var_853, x = var_849_cast_fp16)[name = string("query_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_853 = const()[name = string("op_853"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_849_cast_fp16 = transpose(perm = var_848, x = x_63_cast_fp16)[name = string("transpose_125")];
+            tensor<fp16, [16, 224, 64]> query_layer_21_cast_fp16 = reshape(shape = var_853, x = var_849_cast_fp16)[name = string("query_layer_21_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(347191744)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349288960)))];
-            tensor<fp16, [1, 512, 1024]> linear_25_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_25_cast_fp16")];
-            tensor<int32, [4]> var_860 = const()[name = string("op_860"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_67_cast_fp16 = reshape(shape = var_860, x = linear_25_cast_fp16)[name = string("x_67_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_25_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_key_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_25_cast_fp16")];
+            tensor<int32, [4]> var_860 = const()[name = string("op_860"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_67_cast_fp16 = reshape(shape = var_860, x = linear_25_cast_fp16)[name = string("x_67_cast_fp16")];
             tensor<int32, [4]> var_862 = const()[name = string("op_862"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_867 = const()[name = string("op_867"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_863_cast_fp16 = transpose(perm = var_862, x = x_67_cast_fp16)[name = string("transpose_124")];
-            tensor<fp16, [16, 512, 64]> key_layer_7_cast_fp16 = reshape(shape = var_867, x = var_863_cast_fp16)[name = string("key_layer_7_cast_fp16")];
+            tensor<int32, [3]> var_867 = const()[name = string("op_867"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_863_cast_fp16 = transpose(perm = var_862, x = x_67_cast_fp16)[name = string("transpose_124")];
+            tensor<fp16, [16, 224, 64]> key_layer_7_cast_fp16 = reshape(shape = var_867, x = var_863_cast_fp16)[name = string("key_layer_7_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349291072)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351388288)))];
-            tensor<fp16, [1, 512, 1024]> linear_26_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_26_cast_fp16")];
-            tensor<int32, [4]> var_874 = const()[name = string("op_874"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_71_cast_fp16 = reshape(shape = var_874, x = linear_26_cast_fp16)[name = string("x_71_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_26_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_3_attention_self_value_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_26_cast_fp16")];
+            tensor<int32, [4]> var_874 = const()[name = string("op_874"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_71_cast_fp16 = reshape(shape = var_874, x = linear_26_cast_fp16)[name = string("x_71_cast_fp16")];
             tensor<int32, [4]> var_876 = const()[name = string("op_876"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_881 = const()[name = string("op_881"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_877_cast_fp16 = transpose(perm = var_876, x = x_71_cast_fp16)[name = string("transpose_123")];
-            tensor<fp16, [16, 512, 64]> value_layer_7_cast_fp16 = reshape(shape = var_881, x = var_877_cast_fp16)[name = string("value_layer_7_cast_fp16")];
+            tensor<int32, [3]> var_881 = const()[name = string("op_881"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_877_cast_fp16 = transpose(perm = var_876, x = x_71_cast_fp16)[name = string("transpose_123")];
+            tensor<fp16, [16, 224, 64]> value_layer_7_cast_fp16 = reshape(shape = var_881, x = var_877_cast_fp16)[name = string("value_layer_7_cast_fp16")];
             tensor<int32, [3]> var_891_perm_0 = const()[name = string("op_891_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_893_y_0_to_fp16 = const()[name = string("_inversed_893_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_891_cast_fp16 = transpose(perm = var_891_perm_0, x = key_layer_7_cast_fp16)[name = string("transpose_122")];
-            tensor<fp16, [16, 64, 512]> _inversed_893_cast_fp16 = mul(x = var_891_cast_fp16, y = _inversed_893_y_0_to_fp16)[name = string("_inversed_893_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_891_cast_fp16 = transpose(perm = var_891_perm_0, x = key_layer_7_cast_fp16)[name = string("transpose_122")];
+            tensor<fp16, [16, 64, 224]> _inversed_893_cast_fp16 = mul(x = var_891_cast_fp16, y = _inversed_893_y_0_to_fp16)[name = string("_inversed_893_cast_fp16")];
             bool attention_scores_19_transpose_x_0 = const()[name = string("attention_scores_19_transpose_x_0"), val = bool(false)];
             bool attention_scores_19_transpose_y_0 = const()[name = string("attention_scores_19_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = query_layer_21_cast_fp16, y = _inversed_893_cast_fp16)[name = string("attention_scores_19_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_19_cast_fp16 = matmul(transpose_x = attention_scores_19_transpose_x_0, transpose_y = attention_scores_19_transpose_y_0, x = query_layer_21_cast_fp16, y = _inversed_893_cast_fp16)[name = string("attention_scores_19_cast_fp16")];
             bool c2p_att_13_transpose_x_0 = const()[name = string("c2p_att_13_transpose_x_0"), val = bool(false)];
             bool c2p_att_13_transpose_y_0 = const()[name = string("c2p_att_13_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_949_to_fp16 = const()[name = string("op_949_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(351390400)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_13_cast_fp16 = matmul(transpose_x = c2p_att_13_transpose_x_0, transpose_y = c2p_att_13_transpose_y_0, x = query_layer_21_cast_fp16, y = var_949_to_fp16)[name = string("c2p_att_13_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_13_cast_fp16 = matmul(transpose_x = c2p_att_13_transpose_x_0, transpose_y = c2p_att_13_transpose_y_0, x = query_layer_21_cast_fp16, y = var_949_to_fp16)[name = string("c2p_att_13_cast_fp16")];
             bool c2p_att_15_validate_indices_0 = const()[name = string("c2p_att_15_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_15_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_15_validate_indices_0, x = c2p_att_13_cast_fp16)[name = string("c2p_att_15_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_15_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_15_validate_indices_0, x = c2p_att_13_cast_fp16)[name = string("c2p_att_15_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_962_y_0_to_fp16 = const()[name = string("_inversed_962_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_962_cast_fp16 = mul(x = c2p_att_15_cast_fp16_cast_int16, y = _inversed_962_y_0_to_fp16)[name = string("_inversed_962_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_962_cast_fp16 = mul(x = c2p_att_15_cast_fp16_cast_int16, y = _inversed_962_y_0_to_fp16)[name = string("_inversed_962_cast_fp16")];
             bool p2c_att_13_transpose_x_0 = const()[name = string("p2c_att_13_transpose_x_0"), val = bool(false)];
             bool p2c_att_13_transpose_y_0 = const()[name = string("p2c_att_13_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_977_to_fp16 = const()[name = string("op_977_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352439040)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_13_cast_fp16 = matmul(transpose_x = p2c_att_13_transpose_x_0, transpose_y = p2c_att_13_transpose_y_0, x = key_layer_7_cast_fp16, y = var_977_to_fp16)[name = string("p2c_att_13_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_13_cast_fp16 = matmul(transpose_x = p2c_att_13_transpose_x_0, transpose_y = p2c_att_13_transpose_y_0, x = key_layer_7_cast_fp16, y = var_977_to_fp16)[name = string("p2c_att_13_cast_fp16")];
             bool var_985_validate_indices_0 = const()[name = string("op_985_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_985_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_985_validate_indices_0, x = p2c_att_13_cast_fp16)[name = string("op_985_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_985_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_985_validate_indices_0, x = p2c_att_13_cast_fp16)[name = string("op_985_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_15_perm_0 = const()[name = string("p2c_att_15_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_988_y_0_to_fp16 = const()[name = string("_inversed_988_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_15_cast_fp16 = transpose(perm = p2c_att_15_perm_0, x = var_985_cast_fp16_cast_int16)[name = string("transpose_121")];
-            tensor<fp16, [16, 512, 512]> _inversed_988_cast_fp16 = mul(x = p2c_att_15_cast_fp16, y = _inversed_988_y_0_to_fp16)[name = string("_inversed_988_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_7_cast_fp16 = add(x = _inversed_962_cast_fp16, y = _inversed_988_cast_fp16)[name = string("rel_att_7_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_21_cast_fp16 = add(x = attention_scores_19_cast_fp16, y = rel_att_7_cast_fp16)[name = string("attention_scores_21_cast_fp16")];
-            tensor<int32, [4]> var_993 = const()[name = string("op_993"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_23_cast_fp16 = reshape(shape = var_993, x = attention_scores_21_cast_fp16)[name = string("attention_scores_23_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_81_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_23_cast_fp16, cond = var_339)[name = string("input_81_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_83_cast_fp16 = softmax(axis = var_56, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
-            tensor<int32, [3]> var_1002 = const()[name = string("op_1002"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_1003_cast_fp16 = reshape(shape = var_1002, x = input_83_cast_fp16)[name = string("op_1003_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_15_cast_fp16 = transpose(perm = p2c_att_15_perm_0, x = var_985_cast_fp16_cast_int16)[name = string("transpose_121")];
+            tensor<fp16, [16, 224, 224]> _inversed_988_cast_fp16 = mul(x = p2c_att_15_cast_fp16, y = _inversed_988_y_0_to_fp16)[name = string("_inversed_988_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_7_cast_fp16 = add(x = _inversed_962_cast_fp16, y = _inversed_988_cast_fp16)[name = string("rel_att_7_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_21_cast_fp16 = add(x = attention_scores_19_cast_fp16, y = rel_att_7_cast_fp16)[name = string("attention_scores_21_cast_fp16")];
+            tensor<int32, [4]> var_993 = const()[name = string("op_993"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_23_cast_fp16 = reshape(shape = var_993, x = attention_scores_21_cast_fp16)[name = string("attention_scores_23_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_81_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_23_cast_fp16, cond = var_339)[name = string("input_81_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_83_cast_fp16 = softmax(axis = var_56, x = input_81_cast_fp16)[name = string("input_83_cast_fp16")];
+            tensor<int32, [3]> var_1002 = const()[name = string("op_1002"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_1003_cast_fp16 = reshape(shape = var_1002, x = input_83_cast_fp16)[name = string("op_1003_cast_fp16")];
             bool context_layer_13_transpose_x_0 = const()[name = string("context_layer_13_transpose_x_0"), val = bool(false)];
             bool context_layer_13_transpose_y_0 = const()[name = string("context_layer_13_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = var_1003_cast_fp16, y = value_layer_7_cast_fp16)[name = string("context_layer_13_cast_fp16")];
-            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1008_cast_fp16 = reshape(shape = var_1007, x = context_layer_13_cast_fp16)[name = string("op_1008_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_13_cast_fp16 = matmul(transpose_x = context_layer_13_transpose_x_0, transpose_y = context_layer_13_transpose_y_0, x = var_1003_cast_fp16, y = value_layer_7_cast_fp16)[name = string("context_layer_13_cast_fp16")];
+            tensor<int32, [4]> var_1007 = const()[name = string("op_1007"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1008_cast_fp16 = reshape(shape = var_1007, x = context_layer_13_cast_fp16)[name = string("op_1008_cast_fp16")];
             tensor<int32, [4]> var_1009 = const()[name = string("op_1009"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_1010_cast_fp16 = transpose(perm = var_1009, x = var_1008_cast_fp16)[name = string("transpose_120")];
-            tensor<fp16, [1, 512, 1024]> input_85_cast_fp16 = reshape(shape = var_1014, x = var_1010_cast_fp16)[name = string("input_85_cast_fp16")];
+            tensor<int32, [3]> var_1014 = const()[name = string("op_1014"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_1010_cast_fp16 = transpose(perm = var_1009, x = var_1008_cast_fp16)[name = string("transpose_120")];
+            tensor<fp16, [1, 224, 1024]> input_85_cast_fp16 = reshape(shape = var_1014, x = var_1010_cast_fp16)[name = string("input_85_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_3_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(353487680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355584896)))];
-            tensor<fp16, [1, 512, 1024]> linear_29_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_29_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_89_cast_fp16 = add(x = linear_29_cast_fp16, y = input_77_cast_fp16)[name = string("input_89_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_29_cast_fp16 = linear(bias = encoder_encoder_layer_3_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_attention_output_dense_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_29_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_89_cast_fp16 = add(x = linear_29_cast_fp16, y = input_77_cast_fp16)[name = string("input_89_cast_fp16")];
             tensor<int32, [1]> input_91_axes_0 = const()[name = string("input_91_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355587008)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355589120)))];
-            tensor<fp16, [1, 512, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_91_cast_fp16 = layer_norm(axes = input_91_axes_0, beta = encoder_encoder_layer_3_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_attention_output_LayerNorm_weight_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_3_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355591232)))];
             tensor<fp16, [4096]> encoder_encoder_layer_3_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363979904)))];
-            tensor<fp16, [1, 512, 4096]> linear_30_cast_fp16 = linear(bias = encoder_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_30_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_30_cast_fp16 = linear(bias = encoder_encoder_layer_3_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_3_intermediate_dense_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_30_cast_fp16")];
             string input_95_mode_0 = const()[name = string("input_95_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = string("input_95_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_95_cast_fp16 = gelu(mode = input_95_mode_0, x = linear_30_cast_fp16)[name = string("input_95_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_3_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(363988160)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372376832)))];
-            tensor<fp16, [1, 512, 1024]> linear_31_cast_fp16 = linear(bias = encoder_encoder_layer_3_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_output_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_31_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_99_cast_fp16 = add(x = linear_31_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_31_cast_fp16 = linear(bias = encoder_encoder_layer_3_output_dense_bias_to_fp16, weight = encoder_encoder_layer_3_output_dense_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_31_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_99_cast_fp16 = add(x = linear_31_cast_fp16, y = input_91_cast_fp16)[name = string("input_99_cast_fp16")];
             tensor<int32, [1]> input_101_axes_0 = const()[name = string("input_101_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372378944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372381056)))];
-            tensor<fp16, [1, 512, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_101_cast_fp16 = layer_norm(axes = input_101_axes_0, beta = encoder_encoder_layer_3_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_3_output_LayerNorm_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372383168)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374480384)))];
-            tensor<fp16, [1, 512, 1024]> linear_32_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_32_cast_fp16")];
-            tensor<int32, [4]> var_1065 = const()[name = string("op_1065"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_83_cast_fp16 = reshape(shape = var_1065, x = linear_32_cast_fp16)[name = string("x_83_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_32_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_query_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_32_cast_fp16")];
+            tensor<int32, [4]> var_1065 = const()[name = string("op_1065"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_83_cast_fp16 = reshape(shape = var_1065, x = linear_32_cast_fp16)[name = string("x_83_cast_fp16")];
             tensor<int32, [4]> var_1067 = const()[name = string("op_1067"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1068_cast_fp16 = transpose(perm = var_1067, x = x_83_cast_fp16)[name = string("transpose_119")];
-            tensor<fp16, [16, 512, 64]> query_layer_27_cast_fp16 = reshape(shape = var_1072, x = var_1068_cast_fp16)[name = string("query_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_1072 = const()[name = string("op_1072"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1068_cast_fp16 = transpose(perm = var_1067, x = x_83_cast_fp16)[name = string("transpose_119")];
+            tensor<fp16, [16, 224, 64]> query_layer_27_cast_fp16 = reshape(shape = var_1072, x = var_1068_cast_fp16)[name = string("query_layer_27_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374482496)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376579712)))];
-            tensor<fp16, [1, 512, 1024]> linear_33_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")];
-            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_87_cast_fp16 = reshape(shape = var_1079, x = linear_33_cast_fp16)[name = string("x_87_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_33_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_key_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_33_cast_fp16")];
+            tensor<int32, [4]> var_1079 = const()[name = string("op_1079"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_87_cast_fp16 = reshape(shape = var_1079, x = linear_33_cast_fp16)[name = string("x_87_cast_fp16")];
             tensor<int32, [4]> var_1081 = const()[name = string("op_1081"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1082_cast_fp16 = transpose(perm = var_1081, x = x_87_cast_fp16)[name = string("transpose_118")];
-            tensor<fp16, [16, 512, 64]> key_layer_9_cast_fp16 = reshape(shape = var_1086, x = var_1082_cast_fp16)[name = string("key_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_1086 = const()[name = string("op_1086"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1082_cast_fp16 = transpose(perm = var_1081, x = x_87_cast_fp16)[name = string("transpose_118")];
+            tensor<fp16, [16, 224, 64]> key_layer_9_cast_fp16 = reshape(shape = var_1086, x = var_1082_cast_fp16)[name = string("key_layer_9_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376581824)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378679040)))];
-            tensor<fp16, [1, 512, 1024]> linear_34_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_34_cast_fp16")];
-            tensor<int32, [4]> var_1093 = const()[name = string("op_1093"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_91_cast_fp16 = reshape(shape = var_1093, x = linear_34_cast_fp16)[name = string("x_91_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_34_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_4_attention_self_value_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_34_cast_fp16")];
+            tensor<int32, [4]> var_1093 = const()[name = string("op_1093"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_91_cast_fp16 = reshape(shape = var_1093, x = linear_34_cast_fp16)[name = string("x_91_cast_fp16")];
             tensor<int32, [4]> var_1095 = const()[name = string("op_1095"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1096_cast_fp16 = transpose(perm = var_1095, x = x_91_cast_fp16)[name = string("transpose_117")];
-            tensor<fp16, [16, 512, 64]> value_layer_9_cast_fp16 = reshape(shape = var_1100, x = var_1096_cast_fp16)[name = string("value_layer_9_cast_fp16")];
+            tensor<int32, [3]> var_1100 = const()[name = string("op_1100"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1096_cast_fp16 = transpose(perm = var_1095, x = x_91_cast_fp16)[name = string("transpose_117")];
+            tensor<fp16, [16, 224, 64]> value_layer_9_cast_fp16 = reshape(shape = var_1100, x = var_1096_cast_fp16)[name = string("value_layer_9_cast_fp16")];
             tensor<int32, [3]> var_1110_perm_0 = const()[name = string("op_1110_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1112_y_0_to_fp16 = const()[name = string("_inversed_1112_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_1110_cast_fp16 = transpose(perm = var_1110_perm_0, x = key_layer_9_cast_fp16)[name = string("transpose_116")];
-            tensor<fp16, [16, 64, 512]> _inversed_1112_cast_fp16 = mul(x = var_1110_cast_fp16, y = _inversed_1112_y_0_to_fp16)[name = string("_inversed_1112_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_1110_cast_fp16 = transpose(perm = var_1110_perm_0, x = key_layer_9_cast_fp16)[name = string("transpose_116")];
+            tensor<fp16, [16, 64, 224]> _inversed_1112_cast_fp16 = mul(x = var_1110_cast_fp16, y = _inversed_1112_y_0_to_fp16)[name = string("_inversed_1112_cast_fp16")];
             bool attention_scores_25_transpose_x_0 = const()[name = string("attention_scores_25_transpose_x_0"), val = bool(false)];
             bool attention_scores_25_transpose_y_0 = const()[name = string("attention_scores_25_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_0, transpose_y = attention_scores_25_transpose_y_0, x = query_layer_27_cast_fp16, y = _inversed_1112_cast_fp16)[name = string("attention_scores_25_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_25_cast_fp16 = matmul(transpose_x = attention_scores_25_transpose_x_0, transpose_y = attention_scores_25_transpose_y_0, x = query_layer_27_cast_fp16, y = _inversed_1112_cast_fp16)[name = string("attention_scores_25_cast_fp16")];
             bool c2p_att_17_transpose_x_0 = const()[name = string("c2p_att_17_transpose_x_0"), val = bool(false)];
             bool c2p_att_17_transpose_y_0 = const()[name = string("c2p_att_17_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1168_to_fp16 = const()[name = string("op_1168_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378681152)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_17_cast_fp16 = matmul(transpose_x = c2p_att_17_transpose_x_0, transpose_y = c2p_att_17_transpose_y_0, x = query_layer_27_cast_fp16, y = var_1168_to_fp16)[name = string("c2p_att_17_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_17_cast_fp16 = matmul(transpose_x = c2p_att_17_transpose_x_0, transpose_y = c2p_att_17_transpose_y_0, x = query_layer_27_cast_fp16, y = var_1168_to_fp16)[name = string("c2p_att_17_cast_fp16")];
             bool c2p_att_19_validate_indices_0 = const()[name = string("c2p_att_19_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_19_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_19_validate_indices_0, x = c2p_att_17_cast_fp16)[name = string("c2p_att_19_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_19_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_19_validate_indices_0, x = c2p_att_17_cast_fp16)[name = string("c2p_att_19_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1181_y_0_to_fp16 = const()[name = string("_inversed_1181_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_1181_cast_fp16 = mul(x = c2p_att_19_cast_fp16_cast_int16, y = _inversed_1181_y_0_to_fp16)[name = string("_inversed_1181_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_1181_cast_fp16 = mul(x = c2p_att_19_cast_fp16_cast_int16, y = _inversed_1181_y_0_to_fp16)[name = string("_inversed_1181_cast_fp16")];
             bool p2c_att_17_transpose_x_0 = const()[name = string("p2c_att_17_transpose_x_0"), val = bool(false)];
             bool p2c_att_17_transpose_y_0 = const()[name = string("p2c_att_17_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1196_to_fp16 = const()[name = string("op_1196_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(379729792)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_17_cast_fp16 = matmul(transpose_x = p2c_att_17_transpose_x_0, transpose_y = p2c_att_17_transpose_y_0, x = key_layer_9_cast_fp16, y = var_1196_to_fp16)[name = string("p2c_att_17_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_17_cast_fp16 = matmul(transpose_x = p2c_att_17_transpose_x_0, transpose_y = p2c_att_17_transpose_y_0, x = key_layer_9_cast_fp16, y = var_1196_to_fp16)[name = string("p2c_att_17_cast_fp16")];
             bool var_1204_validate_indices_0 = const()[name = string("op_1204_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_1204_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1204_validate_indices_0, x = p2c_att_17_cast_fp16)[name = string("op_1204_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_1204_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1204_validate_indices_0, x = p2c_att_17_cast_fp16)[name = string("op_1204_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_19_perm_0 = const()[name = string("p2c_att_19_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1207_y_0_to_fp16 = const()[name = string("_inversed_1207_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_19_cast_fp16 = transpose(perm = p2c_att_19_perm_0, x = var_1204_cast_fp16_cast_int16)[name = string("transpose_115")];
-            tensor<fp16, [16, 512, 512]> _inversed_1207_cast_fp16 = mul(x = p2c_att_19_cast_fp16, y = _inversed_1207_y_0_to_fp16)[name = string("_inversed_1207_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_9_cast_fp16 = add(x = _inversed_1181_cast_fp16, y = _inversed_1207_cast_fp16)[name = string("rel_att_9_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_27_cast_fp16 = add(x = attention_scores_25_cast_fp16, y = rel_att_9_cast_fp16)[name = string("attention_scores_27_cast_fp16")];
-            tensor<int32, [4]> var_1212 = const()[name = string("op_1212"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_29_cast_fp16 = reshape(shape = var_1212, x = attention_scores_27_cast_fp16)[name = string("attention_scores_29_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_105_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_29_cast_fp16, cond = var_339)[name = string("input_105_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_107_cast_fp16 = softmax(axis = var_56, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
-            tensor<int32, [3]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_1222_cast_fp16 = reshape(shape = var_1221, x = input_107_cast_fp16)[name = string("op_1222_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_19_cast_fp16 = transpose(perm = p2c_att_19_perm_0, x = var_1204_cast_fp16_cast_int16)[name = string("transpose_115")];
+            tensor<fp16, [16, 224, 224]> _inversed_1207_cast_fp16 = mul(x = p2c_att_19_cast_fp16, y = _inversed_1207_y_0_to_fp16)[name = string("_inversed_1207_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_9_cast_fp16 = add(x = _inversed_1181_cast_fp16, y = _inversed_1207_cast_fp16)[name = string("rel_att_9_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_27_cast_fp16 = add(x = attention_scores_25_cast_fp16, y = rel_att_9_cast_fp16)[name = string("attention_scores_27_cast_fp16")];
+            tensor<int32, [4]> var_1212 = const()[name = string("op_1212"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_29_cast_fp16 = reshape(shape = var_1212, x = attention_scores_27_cast_fp16)[name = string("attention_scores_29_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_105_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_29_cast_fp16, cond = var_339)[name = string("input_105_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_107_cast_fp16 = softmax(axis = var_56, x = input_105_cast_fp16)[name = string("input_107_cast_fp16")];
+            tensor<int32, [3]> var_1221 = const()[name = string("op_1221"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_1222_cast_fp16 = reshape(shape = var_1221, x = input_107_cast_fp16)[name = string("op_1222_cast_fp16")];
             bool context_layer_17_transpose_x_0 = const()[name = string("context_layer_17_transpose_x_0"), val = bool(false)];
             bool context_layer_17_transpose_y_0 = const()[name = string("context_layer_17_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = var_1222_cast_fp16, y = value_layer_9_cast_fp16)[name = string("context_layer_17_cast_fp16")];
-            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1227_cast_fp16 = reshape(shape = var_1226, x = context_layer_17_cast_fp16)[name = string("op_1227_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_17_cast_fp16 = matmul(transpose_x = context_layer_17_transpose_x_0, transpose_y = context_layer_17_transpose_y_0, x = var_1222_cast_fp16, y = value_layer_9_cast_fp16)[name = string("context_layer_17_cast_fp16")];
+            tensor<int32, [4]> var_1226 = const()[name = string("op_1226"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1227_cast_fp16 = reshape(shape = var_1226, x = context_layer_17_cast_fp16)[name = string("op_1227_cast_fp16")];
             tensor<int32, [4]> var_1228 = const()[name = string("op_1228"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_1229_cast_fp16 = transpose(perm = var_1228, x = var_1227_cast_fp16)[name = string("transpose_114")];
-            tensor<fp16, [1, 512, 1024]> input_109_cast_fp16 = reshape(shape = var_1233, x = var_1229_cast_fp16)[name = string("input_109_cast_fp16")];
+            tensor<int32, [3]> var_1233 = const()[name = string("op_1233"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_1229_cast_fp16 = transpose(perm = var_1228, x = var_1227_cast_fp16)[name = string("transpose_114")];
+            tensor<fp16, [1, 224, 1024]> input_109_cast_fp16 = reshape(shape = var_1233, x = var_1229_cast_fp16)[name = string("input_109_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_4_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380778432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382875648)))];
-            tensor<fp16, [1, 512, 1024]> linear_37_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_37_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_113_cast_fp16 = add(x = linear_37_cast_fp16, y = input_101_cast_fp16)[name = string("input_113_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_37_cast_fp16 = linear(bias = encoder_encoder_layer_4_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_attention_output_dense_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_37_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_113_cast_fp16 = add(x = linear_37_cast_fp16, y = input_101_cast_fp16)[name = string("input_113_cast_fp16")];
             tensor<int32, [1]> input_115_axes_0 = const()[name = string("input_115_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382877760)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382879872)))];
-            tensor<fp16, [1, 512, 1024]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_115_cast_fp16 = layer_norm(axes = input_115_axes_0, beta = encoder_encoder_layer_4_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_attention_output_LayerNorm_weight_to_fp16, x = input_113_cast_fp16)[name = string("input_115_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_4_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(382881984)))];
             tensor<fp16, [4096]> encoder_encoder_layer_4_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391270656)))];
-            tensor<fp16, [1, 512, 4096]> linear_38_cast_fp16 = linear(bias = encoder_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_38_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_38_cast_fp16 = linear(bias = encoder_encoder_layer_4_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_4_intermediate_dense_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_38_cast_fp16")];
             string input_119_mode_0 = const()[name = string("input_119_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = string("input_119_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_119_cast_fp16 = gelu(mode = input_119_mode_0, x = linear_38_cast_fp16)[name = string("input_119_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_4_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(391278912)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399667584)))];
-            tensor<fp16, [1, 512, 1024]> linear_39_cast_fp16 = linear(bias = encoder_encoder_layer_4_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_output_dense_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_39_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_123_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = string("input_123_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_39_cast_fp16 = linear(bias = encoder_encoder_layer_4_output_dense_bias_to_fp16, weight = encoder_encoder_layer_4_output_dense_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_39_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_123_cast_fp16 = add(x = linear_39_cast_fp16, y = input_115_cast_fp16)[name = string("input_123_cast_fp16")];
             tensor<int32, [1]> input_125_axes_0 = const()[name = string("input_125_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399669696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399671808)))];
-            tensor<fp16, [1, 512, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_125_cast_fp16 = layer_norm(axes = input_125_axes_0, beta = encoder_encoder_layer_4_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_4_output_LayerNorm_weight_to_fp16, x = input_123_cast_fp16)[name = string("input_125_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399673920)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401771136)))];
-            tensor<fp16, [1, 512, 1024]> linear_40_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_40_cast_fp16")];
-            tensor<int32, [4]> var_1284 = const()[name = string("op_1284"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_103_cast_fp16 = reshape(shape = var_1284, x = linear_40_cast_fp16)[name = string("x_103_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_40_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_query_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_40_cast_fp16")];
+            tensor<int32, [4]> var_1284 = const()[name = string("op_1284"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_103_cast_fp16 = reshape(shape = var_1284, x = linear_40_cast_fp16)[name = string("x_103_cast_fp16")];
             tensor<int32, [4]> var_1286 = const()[name = string("op_1286"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1291 = const()[name = string("op_1291"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1287_cast_fp16 = transpose(perm = var_1286, x = x_103_cast_fp16)[name = string("transpose_113")];
-            tensor<fp16, [16, 512, 64]> query_layer_33_cast_fp16 = reshape(shape = var_1291, x = var_1287_cast_fp16)[name = string("query_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_1291 = const()[name = string("op_1291"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1287_cast_fp16 = transpose(perm = var_1286, x = x_103_cast_fp16)[name = string("transpose_113")];
+            tensor<fp16, [16, 224, 64]> query_layer_33_cast_fp16 = reshape(shape = var_1291, x = var_1287_cast_fp16)[name = string("query_layer_33_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(401773248)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403870464)))];
-            tensor<fp16, [1, 512, 1024]> linear_41_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_41_cast_fp16")];
-            tensor<int32, [4]> var_1298 = const()[name = string("op_1298"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_107_cast_fp16 = reshape(shape = var_1298, x = linear_41_cast_fp16)[name = string("x_107_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_41_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_key_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_41_cast_fp16")];
+            tensor<int32, [4]> var_1298 = const()[name = string("op_1298"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_107_cast_fp16 = reshape(shape = var_1298, x = linear_41_cast_fp16)[name = string("x_107_cast_fp16")];
             tensor<int32, [4]> var_1300 = const()[name = string("op_1300"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1305 = const()[name = string("op_1305"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1301_cast_fp16 = transpose(perm = var_1300, x = x_107_cast_fp16)[name = string("transpose_112")];
-            tensor<fp16, [16, 512, 64]> key_layer_11_cast_fp16 = reshape(shape = var_1305, x = var_1301_cast_fp16)[name = string("key_layer_11_cast_fp16")];
+            tensor<int32, [3]> var_1305 = const()[name = string("op_1305"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1301_cast_fp16 = transpose(perm = var_1300, x = x_107_cast_fp16)[name = string("transpose_112")];
+            tensor<fp16, [16, 224, 64]> key_layer_11_cast_fp16 = reshape(shape = var_1305, x = var_1301_cast_fp16)[name = string("key_layer_11_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(403872576)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405969792)))];
-            tensor<fp16, [1, 512, 1024]> linear_42_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_42_cast_fp16")];
-            tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_111_cast_fp16 = reshape(shape = var_1312, x = linear_42_cast_fp16)[name = string("x_111_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_42_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_5_attention_self_value_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_42_cast_fp16")];
+            tensor<int32, [4]> var_1312 = const()[name = string("op_1312"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_111_cast_fp16 = reshape(shape = var_1312, x = linear_42_cast_fp16)[name = string("x_111_cast_fp16")];
             tensor<int32, [4]> var_1314 = const()[name = string("op_1314"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1319 = const()[name = string("op_1319"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1315_cast_fp16 = transpose(perm = var_1314, x = x_111_cast_fp16)[name = string("transpose_111")];
-            tensor<fp16, [16, 512, 64]> value_layer_11_cast_fp16 = reshape(shape = var_1319, x = var_1315_cast_fp16)[name = string("value_layer_11_cast_fp16")];
+            tensor<int32, [3]> var_1319 = const()[name = string("op_1319"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1315_cast_fp16 = transpose(perm = var_1314, x = x_111_cast_fp16)[name = string("transpose_111")];
+            tensor<fp16, [16, 224, 64]> value_layer_11_cast_fp16 = reshape(shape = var_1319, x = var_1315_cast_fp16)[name = string("value_layer_11_cast_fp16")];
             tensor<int32, [3]> var_1329_perm_0 = const()[name = string("op_1329_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1331_y_0_to_fp16 = const()[name = string("_inversed_1331_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_1329_cast_fp16 = transpose(perm = var_1329_perm_0, x = key_layer_11_cast_fp16)[name = string("transpose_110")];
-            tensor<fp16, [16, 64, 512]> _inversed_1331_cast_fp16 = mul(x = var_1329_cast_fp16, y = _inversed_1331_y_0_to_fp16)[name = string("_inversed_1331_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_1329_cast_fp16 = transpose(perm = var_1329_perm_0, x = key_layer_11_cast_fp16)[name = string("transpose_110")];
+            tensor<fp16, [16, 64, 224]> _inversed_1331_cast_fp16 = mul(x = var_1329_cast_fp16, y = _inversed_1331_y_0_to_fp16)[name = string("_inversed_1331_cast_fp16")];
             bool attention_scores_31_transpose_x_0 = const()[name = string("attention_scores_31_transpose_x_0"), val = bool(false)];
             bool attention_scores_31_transpose_y_0 = const()[name = string("attention_scores_31_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_31_cast_fp16 = matmul(transpose_x = attention_scores_31_transpose_x_0, transpose_y = attention_scores_31_transpose_y_0, x = query_layer_33_cast_fp16, y = _inversed_1331_cast_fp16)[name = string("attention_scores_31_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_31_cast_fp16 = matmul(transpose_x = attention_scores_31_transpose_x_0, transpose_y = attention_scores_31_transpose_y_0, x = query_layer_33_cast_fp16, y = _inversed_1331_cast_fp16)[name = string("attention_scores_31_cast_fp16")];
             bool c2p_att_21_transpose_x_0 = const()[name = string("c2p_att_21_transpose_x_0"), val = bool(false)];
             bool c2p_att_21_transpose_y_0 = const()[name = string("c2p_att_21_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1387_to_fp16 = const()[name = string("op_1387_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405971904)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_21_cast_fp16 = matmul(transpose_x = c2p_att_21_transpose_x_0, transpose_y = c2p_att_21_transpose_y_0, x = query_layer_33_cast_fp16, y = var_1387_to_fp16)[name = string("c2p_att_21_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_21_cast_fp16 = matmul(transpose_x = c2p_att_21_transpose_x_0, transpose_y = c2p_att_21_transpose_y_0, x = query_layer_33_cast_fp16, y = var_1387_to_fp16)[name = string("c2p_att_21_cast_fp16")];
             bool c2p_att_23_validate_indices_0 = const()[name = string("c2p_att_23_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_23_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_23_validate_indices_0, x = c2p_att_21_cast_fp16)[name = string("c2p_att_23_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_23_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_23_validate_indices_0, x = c2p_att_21_cast_fp16)[name = string("c2p_att_23_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1400_y_0_to_fp16 = const()[name = string("_inversed_1400_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_1400_cast_fp16 = mul(x = c2p_att_23_cast_fp16_cast_int16, y = _inversed_1400_y_0_to_fp16)[name = string("_inversed_1400_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_1400_cast_fp16 = mul(x = c2p_att_23_cast_fp16_cast_int16, y = _inversed_1400_y_0_to_fp16)[name = string("_inversed_1400_cast_fp16")];
             bool p2c_att_21_transpose_x_0 = const()[name = string("p2c_att_21_transpose_x_0"), val = bool(false)];
             bool p2c_att_21_transpose_y_0 = const()[name = string("p2c_att_21_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1415_to_fp16 = const()[name = string("op_1415_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407020544)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_21_cast_fp16 = matmul(transpose_x = p2c_att_21_transpose_x_0, transpose_y = p2c_att_21_transpose_y_0, x = key_layer_11_cast_fp16, y = var_1415_to_fp16)[name = string("p2c_att_21_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_21_cast_fp16 = matmul(transpose_x = p2c_att_21_transpose_x_0, transpose_y = p2c_att_21_transpose_y_0, x = key_layer_11_cast_fp16, y = var_1415_to_fp16)[name = string("p2c_att_21_cast_fp16")];
             bool var_1423_validate_indices_0 = const()[name = string("op_1423_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_1423_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1423_validate_indices_0, x = p2c_att_21_cast_fp16)[name = string("op_1423_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_1423_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1423_validate_indices_0, x = p2c_att_21_cast_fp16)[name = string("op_1423_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_23_perm_0 = const()[name = string("p2c_att_23_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1426_y_0_to_fp16 = const()[name = string("_inversed_1426_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_23_cast_fp16 = transpose(perm = p2c_att_23_perm_0, x = var_1423_cast_fp16_cast_int16)[name = string("transpose_109")];
-            tensor<fp16, [16, 512, 512]> _inversed_1426_cast_fp16 = mul(x = p2c_att_23_cast_fp16, y = _inversed_1426_y_0_to_fp16)[name = string("_inversed_1426_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_11_cast_fp16 = add(x = _inversed_1400_cast_fp16, y = _inversed_1426_cast_fp16)[name = string("rel_att_11_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_33_cast_fp16 = add(x = attention_scores_31_cast_fp16, y = rel_att_11_cast_fp16)[name = string("attention_scores_33_cast_fp16")];
-            tensor<int32, [4]> var_1431 = const()[name = string("op_1431"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_35_cast_fp16 = reshape(shape = var_1431, x = attention_scores_33_cast_fp16)[name = string("attention_scores_35_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_129_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_35_cast_fp16, cond = var_339)[name = string("input_129_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_131_cast_fp16 = softmax(axis = var_56, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
-            tensor<int32, [3]> var_1440 = const()[name = string("op_1440"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_1441_cast_fp16 = reshape(shape = var_1440, x = input_131_cast_fp16)[name = string("op_1441_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_23_cast_fp16 = transpose(perm = p2c_att_23_perm_0, x = var_1423_cast_fp16_cast_int16)[name = string("transpose_109")];
+            tensor<fp16, [16, 224, 224]> _inversed_1426_cast_fp16 = mul(x = p2c_att_23_cast_fp16, y = _inversed_1426_y_0_to_fp16)[name = string("_inversed_1426_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_11_cast_fp16 = add(x = _inversed_1400_cast_fp16, y = _inversed_1426_cast_fp16)[name = string("rel_att_11_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_33_cast_fp16 = add(x = attention_scores_31_cast_fp16, y = rel_att_11_cast_fp16)[name = string("attention_scores_33_cast_fp16")];
+            tensor<int32, [4]> var_1431 = const()[name = string("op_1431"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_35_cast_fp16 = reshape(shape = var_1431, x = attention_scores_33_cast_fp16)[name = string("attention_scores_35_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_129_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_35_cast_fp16, cond = var_339)[name = string("input_129_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_131_cast_fp16 = softmax(axis = var_56, x = input_129_cast_fp16)[name = string("input_131_cast_fp16")];
+            tensor<int32, [3]> var_1440 = const()[name = string("op_1440"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_1441_cast_fp16 = reshape(shape = var_1440, x = input_131_cast_fp16)[name = string("op_1441_cast_fp16")];
             bool context_layer_21_transpose_x_0 = const()[name = string("context_layer_21_transpose_x_0"), val = bool(false)];
             bool context_layer_21_transpose_y_0 = const()[name = string("context_layer_21_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = var_1441_cast_fp16, y = value_layer_11_cast_fp16)[name = string("context_layer_21_cast_fp16")];
-            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1446_cast_fp16 = reshape(shape = var_1445, x = context_layer_21_cast_fp16)[name = string("op_1446_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_21_cast_fp16 = matmul(transpose_x = context_layer_21_transpose_x_0, transpose_y = context_layer_21_transpose_y_0, x = var_1441_cast_fp16, y = value_layer_11_cast_fp16)[name = string("context_layer_21_cast_fp16")];
+            tensor<int32, [4]> var_1445 = const()[name = string("op_1445"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1446_cast_fp16 = reshape(shape = var_1445, x = context_layer_21_cast_fp16)[name = string("op_1446_cast_fp16")];
             tensor<int32, [4]> var_1447 = const()[name = string("op_1447"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_1448_cast_fp16 = transpose(perm = var_1447, x = var_1446_cast_fp16)[name = string("transpose_108")];
-            tensor<fp16, [1, 512, 1024]> input_133_cast_fp16 = reshape(shape = var_1452, x = var_1448_cast_fp16)[name = string("input_133_cast_fp16")];
+            tensor<int32, [3]> var_1452 = const()[name = string("op_1452"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_1448_cast_fp16 = transpose(perm = var_1447, x = var_1446_cast_fp16)[name = string("transpose_108")];
+            tensor<fp16, [1, 224, 1024]> input_133_cast_fp16 = reshape(shape = var_1452, x = var_1448_cast_fp16)[name = string("input_133_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_5_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408069184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410166400)))];
-            tensor<fp16, [1, 512, 1024]> linear_45_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_45_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_125_cast_fp16)[name = string("input_137_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_45_cast_fp16 = linear(bias = encoder_encoder_layer_5_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_attention_output_dense_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_45_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_137_cast_fp16 = add(x = linear_45_cast_fp16, y = input_125_cast_fp16)[name = string("input_137_cast_fp16")];
             tensor<int32, [1]> input_139_axes_0 = const()[name = string("input_139_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410168512)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410170624)))];
-            tensor<fp16, [1, 512, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_139_cast_fp16 = layer_norm(axes = input_139_axes_0, beta = encoder_encoder_layer_5_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_attention_output_LayerNorm_weight_to_fp16, x = input_137_cast_fp16)[name = string("input_139_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_5_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410172736)))];
             tensor<fp16, [4096]> encoder_encoder_layer_5_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418561408)))];
-            tensor<fp16, [1, 512, 4096]> linear_46_cast_fp16 = linear(bias = encoder_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_46_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_46_cast_fp16 = linear(bias = encoder_encoder_layer_5_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_5_intermediate_dense_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_46_cast_fp16")];
             string input_143_mode_0 = const()[name = string("input_143_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = string("input_143_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_143_cast_fp16 = gelu(mode = input_143_mode_0, x = linear_46_cast_fp16)[name = string("input_143_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_5_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418569664)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426958336)))];
-            tensor<fp16, [1, 512, 1024]> linear_47_cast_fp16 = linear(bias = encoder_encoder_layer_5_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_output_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_47_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_139_cast_fp16)[name = string("input_147_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_47_cast_fp16 = linear(bias = encoder_encoder_layer_5_output_dense_bias_to_fp16, weight = encoder_encoder_layer_5_output_dense_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_47_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_147_cast_fp16 = add(x = linear_47_cast_fp16, y = input_139_cast_fp16)[name = string("input_147_cast_fp16")];
             tensor<int32, [1]> input_149_axes_0 = const()[name = string("input_149_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426960448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426962560)))];
-            tensor<fp16, [1, 512, 1024]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_149_cast_fp16 = layer_norm(axes = input_149_axes_0, beta = encoder_encoder_layer_5_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_5_output_LayerNorm_weight_to_fp16, x = input_147_cast_fp16)[name = string("input_149_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(426964672)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429061888)))];
-            tensor<fp16, [1, 512, 1024]> linear_48_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_48_cast_fp16")];
-            tensor<int32, [4]> var_1503 = const()[name = string("op_1503"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_123_cast_fp16 = reshape(shape = var_1503, x = linear_48_cast_fp16)[name = string("x_123_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_48_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_query_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_48_cast_fp16")];
+            tensor<int32, [4]> var_1503 = const()[name = string("op_1503"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_123_cast_fp16 = reshape(shape = var_1503, x = linear_48_cast_fp16)[name = string("x_123_cast_fp16")];
             tensor<int32, [4]> var_1505 = const()[name = string("op_1505"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1506_cast_fp16 = transpose(perm = var_1505, x = x_123_cast_fp16)[name = string("transpose_107")];
-            tensor<fp16, [16, 512, 64]> query_layer_39_cast_fp16 = reshape(shape = var_1510, x = var_1506_cast_fp16)[name = string("query_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1506_cast_fp16 = transpose(perm = var_1505, x = x_123_cast_fp16)[name = string("transpose_107")];
+            tensor<fp16, [16, 224, 64]> query_layer_39_cast_fp16 = reshape(shape = var_1510, x = var_1506_cast_fp16)[name = string("query_layer_39_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429064000)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431161216)))];
-            tensor<fp16, [1, 512, 1024]> linear_49_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_49_cast_fp16")];
-            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_127_cast_fp16 = reshape(shape = var_1517, x = linear_49_cast_fp16)[name = string("x_127_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_49_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_key_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_49_cast_fp16")];
+            tensor<int32, [4]> var_1517 = const()[name = string("op_1517"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_127_cast_fp16 = reshape(shape = var_1517, x = linear_49_cast_fp16)[name = string("x_127_cast_fp16")];
             tensor<int32, [4]> var_1519 = const()[name = string("op_1519"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1520_cast_fp16 = transpose(perm = var_1519, x = x_127_cast_fp16)[name = string("transpose_106")];
-            tensor<fp16, [16, 512, 64]> key_layer_13_cast_fp16 = reshape(shape = var_1524, x = var_1520_cast_fp16)[name = string("key_layer_13_cast_fp16")];
+            tensor<int32, [3]> var_1524 = const()[name = string("op_1524"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1520_cast_fp16 = transpose(perm = var_1519, x = x_127_cast_fp16)[name = string("transpose_106")];
+            tensor<fp16, [16, 224, 64]> key_layer_13_cast_fp16 = reshape(shape = var_1524, x = var_1520_cast_fp16)[name = string("key_layer_13_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431163328)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433260544)))];
-            tensor<fp16, [1, 512, 1024]> linear_50_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_50_cast_fp16")];
-            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_131_cast_fp16 = reshape(shape = var_1531, x = linear_50_cast_fp16)[name = string("x_131_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_50_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_6_attention_self_value_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_50_cast_fp16")];
+            tensor<int32, [4]> var_1531 = const()[name = string("op_1531"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_131_cast_fp16 = reshape(shape = var_1531, x = linear_50_cast_fp16)[name = string("x_131_cast_fp16")];
             tensor<int32, [4]> var_1533 = const()[name = string("op_1533"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1538 = const()[name = string("op_1538"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1534_cast_fp16 = transpose(perm = var_1533, x = x_131_cast_fp16)[name = string("transpose_105")];
-            tensor<fp16, [16, 512, 64]> value_layer_13_cast_fp16 = reshape(shape = var_1538, x = var_1534_cast_fp16)[name = string("value_layer_13_cast_fp16")];
+            tensor<int32, [3]> var_1538 = const()[name = string("op_1538"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1534_cast_fp16 = transpose(perm = var_1533, x = x_131_cast_fp16)[name = string("transpose_105")];
+            tensor<fp16, [16, 224, 64]> value_layer_13_cast_fp16 = reshape(shape = var_1538, x = var_1534_cast_fp16)[name = string("value_layer_13_cast_fp16")];
             tensor<int32, [3]> var_1548_perm_0 = const()[name = string("op_1548_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1550_y_0_to_fp16 = const()[name = string("_inversed_1550_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_1548_cast_fp16 = transpose(perm = var_1548_perm_0, x = key_layer_13_cast_fp16)[name = string("transpose_104")];
-            tensor<fp16, [16, 64, 512]> _inversed_1550_cast_fp16 = mul(x = var_1548_cast_fp16, y = _inversed_1550_y_0_to_fp16)[name = string("_inversed_1550_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_1548_cast_fp16 = transpose(perm = var_1548_perm_0, x = key_layer_13_cast_fp16)[name = string("transpose_104")];
+            tensor<fp16, [16, 64, 224]> _inversed_1550_cast_fp16 = mul(x = var_1548_cast_fp16, y = _inversed_1550_y_0_to_fp16)[name = string("_inversed_1550_cast_fp16")];
             bool attention_scores_37_transpose_x_0 = const()[name = string("attention_scores_37_transpose_x_0"), val = bool(false)];
             bool attention_scores_37_transpose_y_0 = const()[name = string("attention_scores_37_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_37_cast_fp16 = matmul(transpose_x = attention_scores_37_transpose_x_0, transpose_y = attention_scores_37_transpose_y_0, x = query_layer_39_cast_fp16, y = _inversed_1550_cast_fp16)[name = string("attention_scores_37_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_37_cast_fp16 = matmul(transpose_x = attention_scores_37_transpose_x_0, transpose_y = attention_scores_37_transpose_y_0, x = query_layer_39_cast_fp16, y = _inversed_1550_cast_fp16)[name = string("attention_scores_37_cast_fp16")];
             bool c2p_att_25_transpose_x_0 = const()[name = string("c2p_att_25_transpose_x_0"), val = bool(false)];
             bool c2p_att_25_transpose_y_0 = const()[name = string("c2p_att_25_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1606_to_fp16 = const()[name = string("op_1606_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(433262656)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_25_cast_fp16 = matmul(transpose_x = c2p_att_25_transpose_x_0, transpose_y = c2p_att_25_transpose_y_0, x = query_layer_39_cast_fp16, y = var_1606_to_fp16)[name = string("c2p_att_25_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_25_cast_fp16 = matmul(transpose_x = c2p_att_25_transpose_x_0, transpose_y = c2p_att_25_transpose_y_0, x = query_layer_39_cast_fp16, y = var_1606_to_fp16)[name = string("c2p_att_25_cast_fp16")];
             bool c2p_att_27_validate_indices_0 = const()[name = string("c2p_att_27_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_27_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_27_validate_indices_0, x = c2p_att_25_cast_fp16)[name = string("c2p_att_27_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_27_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_27_validate_indices_0, x = c2p_att_25_cast_fp16)[name = string("c2p_att_27_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1619_y_0_to_fp16 = const()[name = string("_inversed_1619_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_1619_cast_fp16 = mul(x = c2p_att_27_cast_fp16_cast_int16, y = _inversed_1619_y_0_to_fp16)[name = string("_inversed_1619_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_1619_cast_fp16 = mul(x = c2p_att_27_cast_fp16_cast_int16, y = _inversed_1619_y_0_to_fp16)[name = string("_inversed_1619_cast_fp16")];
             bool p2c_att_25_transpose_x_0 = const()[name = string("p2c_att_25_transpose_x_0"), val = bool(false)];
             bool p2c_att_25_transpose_y_0 = const()[name = string("p2c_att_25_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1634_to_fp16 = const()[name = string("op_1634_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434311296)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_25_cast_fp16 = matmul(transpose_x = p2c_att_25_transpose_x_0, transpose_y = p2c_att_25_transpose_y_0, x = key_layer_13_cast_fp16, y = var_1634_to_fp16)[name = string("p2c_att_25_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_25_cast_fp16 = matmul(transpose_x = p2c_att_25_transpose_x_0, transpose_y = p2c_att_25_transpose_y_0, x = key_layer_13_cast_fp16, y = var_1634_to_fp16)[name = string("p2c_att_25_cast_fp16")];
             bool var_1642_validate_indices_0 = const()[name = string("op_1642_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_1642_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1642_validate_indices_0, x = p2c_att_25_cast_fp16)[name = string("op_1642_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_1642_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1642_validate_indices_0, x = p2c_att_25_cast_fp16)[name = string("op_1642_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_27_perm_0 = const()[name = string("p2c_att_27_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1645_y_0_to_fp16 = const()[name = string("_inversed_1645_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_27_cast_fp16 = transpose(perm = p2c_att_27_perm_0, x = var_1642_cast_fp16_cast_int16)[name = string("transpose_103")];
-            tensor<fp16, [16, 512, 512]> _inversed_1645_cast_fp16 = mul(x = p2c_att_27_cast_fp16, y = _inversed_1645_y_0_to_fp16)[name = string("_inversed_1645_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_13_cast_fp16 = add(x = _inversed_1619_cast_fp16, y = _inversed_1645_cast_fp16)[name = string("rel_att_13_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_39_cast_fp16 = add(x = attention_scores_37_cast_fp16, y = rel_att_13_cast_fp16)[name = string("attention_scores_39_cast_fp16")];
-            tensor<int32, [4]> var_1650 = const()[name = string("op_1650"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_41_cast_fp16 = reshape(shape = var_1650, x = attention_scores_39_cast_fp16)[name = string("attention_scores_41_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_153_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_41_cast_fp16, cond = var_339)[name = string("input_153_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_155_cast_fp16 = softmax(axis = var_56, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
-            tensor<int32, [3]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_1660_cast_fp16 = reshape(shape = var_1659, x = input_155_cast_fp16)[name = string("op_1660_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_27_cast_fp16 = transpose(perm = p2c_att_27_perm_0, x = var_1642_cast_fp16_cast_int16)[name = string("transpose_103")];
+            tensor<fp16, [16, 224, 224]> _inversed_1645_cast_fp16 = mul(x = p2c_att_27_cast_fp16, y = _inversed_1645_y_0_to_fp16)[name = string("_inversed_1645_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_13_cast_fp16 = add(x = _inversed_1619_cast_fp16, y = _inversed_1645_cast_fp16)[name = string("rel_att_13_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_39_cast_fp16 = add(x = attention_scores_37_cast_fp16, y = rel_att_13_cast_fp16)[name = string("attention_scores_39_cast_fp16")];
+            tensor<int32, [4]> var_1650 = const()[name = string("op_1650"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_41_cast_fp16 = reshape(shape = var_1650, x = attention_scores_39_cast_fp16)[name = string("attention_scores_41_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_153_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_41_cast_fp16, cond = var_339)[name = string("input_153_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_155_cast_fp16 = softmax(axis = var_56, x = input_153_cast_fp16)[name = string("input_155_cast_fp16")];
+            tensor<int32, [3]> var_1659 = const()[name = string("op_1659"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_1660_cast_fp16 = reshape(shape = var_1659, x = input_155_cast_fp16)[name = string("op_1660_cast_fp16")];
             bool context_layer_25_transpose_x_0 = const()[name = string("context_layer_25_transpose_x_0"), val = bool(false)];
             bool context_layer_25_transpose_y_0 = const()[name = string("context_layer_25_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = var_1660_cast_fp16, y = value_layer_13_cast_fp16)[name = string("context_layer_25_cast_fp16")];
-            tensor<int32, [4]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1665_cast_fp16 = reshape(shape = var_1664, x = context_layer_25_cast_fp16)[name = string("op_1665_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_25_cast_fp16 = matmul(transpose_x = context_layer_25_transpose_x_0, transpose_y = context_layer_25_transpose_y_0, x = var_1660_cast_fp16, y = value_layer_13_cast_fp16)[name = string("context_layer_25_cast_fp16")];
+            tensor<int32, [4]> var_1664 = const()[name = string("op_1664"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1665_cast_fp16 = reshape(shape = var_1664, x = context_layer_25_cast_fp16)[name = string("op_1665_cast_fp16")];
             tensor<int32, [4]> var_1666 = const()[name = string("op_1666"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1671 = const()[name = string("op_1671"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1666, x = var_1665_cast_fp16)[name = string("transpose_102")];
-            tensor<fp16, [1, 512, 1024]> input_157_cast_fp16 = reshape(shape = var_1671, x = var_1667_cast_fp16)[name = string("input_157_cast_fp16")];
+            tensor<int32, [3]> var_1671 = const()[name = string("op_1671"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_1667_cast_fp16 = transpose(perm = var_1666, x = var_1665_cast_fp16)[name = string("transpose_102")];
+            tensor<fp16, [1, 224, 1024]> input_157_cast_fp16 = reshape(shape = var_1671, x = var_1667_cast_fp16)[name = string("input_157_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_6_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(435359936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437457152)))];
-            tensor<fp16, [1, 512, 1024]> linear_53_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_53_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_161_cast_fp16 = add(x = linear_53_cast_fp16, y = input_149_cast_fp16)[name = string("input_161_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_53_cast_fp16 = linear(bias = encoder_encoder_layer_6_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_attention_output_dense_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_53_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_161_cast_fp16 = add(x = linear_53_cast_fp16, y = input_149_cast_fp16)[name = string("input_161_cast_fp16")];
             tensor<int32, [1]> input_163_axes_0 = const()[name = string("input_163_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437459264)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437461376)))];
-            tensor<fp16, [1, 512, 1024]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_163_cast_fp16 = layer_norm(axes = input_163_axes_0, beta = encoder_encoder_layer_6_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_attention_output_LayerNorm_weight_to_fp16, x = input_161_cast_fp16)[name = string("input_163_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_6_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437463488)))];
             tensor<fp16, [4096]> encoder_encoder_layer_6_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445852160)))];
-            tensor<fp16, [1, 512, 4096]> linear_54_cast_fp16 = linear(bias = encoder_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_54_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_54_cast_fp16 = linear(bias = encoder_encoder_layer_6_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_6_intermediate_dense_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_54_cast_fp16")];
             string input_167_mode_0 = const()[name = string("input_167_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = string("input_167_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_167_cast_fp16 = gelu(mode = input_167_mode_0, x = linear_54_cast_fp16)[name = string("input_167_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_6_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(445860416)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454249088)))];
-            tensor<fp16, [1, 512, 1024]> linear_55_cast_fp16 = linear(bias = encoder_encoder_layer_6_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_output_dense_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_55_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_171_cast_fp16 = add(x = linear_55_cast_fp16, y = input_163_cast_fp16)[name = string("input_171_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_55_cast_fp16 = linear(bias = encoder_encoder_layer_6_output_dense_bias_to_fp16, weight = encoder_encoder_layer_6_output_dense_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_55_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_171_cast_fp16 = add(x = linear_55_cast_fp16, y = input_163_cast_fp16)[name = string("input_171_cast_fp16")];
             tensor<int32, [1]> input_173_axes_0 = const()[name = string("input_173_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454251200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454253312)))];
-            tensor<fp16, [1, 512, 1024]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_173_cast_fp16 = layer_norm(axes = input_173_axes_0, beta = encoder_encoder_layer_6_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_6_output_LayerNorm_weight_to_fp16, x = input_171_cast_fp16)[name = string("input_173_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(454255424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456352640)))];
-            tensor<fp16, [1, 512, 1024]> linear_56_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_56_cast_fp16")];
-            tensor<int32, [4]> var_1722 = const()[name = string("op_1722"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_143_cast_fp16 = reshape(shape = var_1722, x = linear_56_cast_fp16)[name = string("x_143_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_56_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_query_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_56_cast_fp16")];
+            tensor<int32, [4]> var_1722 = const()[name = string("op_1722"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_143_cast_fp16 = reshape(shape = var_1722, x = linear_56_cast_fp16)[name = string("x_143_cast_fp16")];
             tensor<int32, [4]> var_1724 = const()[name = string("op_1724"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1729 = const()[name = string("op_1729"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1725_cast_fp16 = transpose(perm = var_1724, x = x_143_cast_fp16)[name = string("transpose_101")];
-            tensor<fp16, [16, 512, 64]> query_layer_45_cast_fp16 = reshape(shape = var_1729, x = var_1725_cast_fp16)[name = string("query_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_1729 = const()[name = string("op_1729"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1725_cast_fp16 = transpose(perm = var_1724, x = x_143_cast_fp16)[name = string("transpose_101")];
+            tensor<fp16, [16, 224, 64]> query_layer_45_cast_fp16 = reshape(shape = var_1729, x = var_1725_cast_fp16)[name = string("query_layer_45_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456354752)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458451968)))];
-            tensor<fp16, [1, 512, 1024]> linear_57_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_57_cast_fp16")];
-            tensor<int32, [4]> var_1736 = const()[name = string("op_1736"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_147_cast_fp16 = reshape(shape = var_1736, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_57_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_key_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_57_cast_fp16")];
+            tensor<int32, [4]> var_1736 = const()[name = string("op_1736"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_147_cast_fp16 = reshape(shape = var_1736, x = linear_57_cast_fp16)[name = string("x_147_cast_fp16")];
             tensor<int32, [4]> var_1738 = const()[name = string("op_1738"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1739_cast_fp16 = transpose(perm = var_1738, x = x_147_cast_fp16)[name = string("transpose_100")];
-            tensor<fp16, [16, 512, 64]> key_layer_15_cast_fp16 = reshape(shape = var_1743, x = var_1739_cast_fp16)[name = string("key_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_1743 = const()[name = string("op_1743"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1739_cast_fp16 = transpose(perm = var_1738, x = x_147_cast_fp16)[name = string("transpose_100")];
+            tensor<fp16, [16, 224, 64]> key_layer_15_cast_fp16 = reshape(shape = var_1743, x = var_1739_cast_fp16)[name = string("key_layer_15_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(458454080)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460551296)))];
-            tensor<fp16, [1, 512, 1024]> linear_58_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_58_cast_fp16")];
-            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_151_cast_fp16 = reshape(shape = var_1750, x = linear_58_cast_fp16)[name = string("x_151_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_58_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_7_attention_self_value_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_58_cast_fp16")];
+            tensor<int32, [4]> var_1750 = const()[name = string("op_1750"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_151_cast_fp16 = reshape(shape = var_1750, x = linear_58_cast_fp16)[name = string("x_151_cast_fp16")];
             tensor<int32, [4]> var_1752 = const()[name = string("op_1752"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1757 = const()[name = string("op_1757"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1753_cast_fp16 = transpose(perm = var_1752, x = x_151_cast_fp16)[name = string("transpose_99")];
-            tensor<fp16, [16, 512, 64]> value_layer_15_cast_fp16 = reshape(shape = var_1757, x = var_1753_cast_fp16)[name = string("value_layer_15_cast_fp16")];
+            tensor<int32, [3]> var_1757 = const()[name = string("op_1757"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1753_cast_fp16 = transpose(perm = var_1752, x = x_151_cast_fp16)[name = string("transpose_99")];
+            tensor<fp16, [16, 224, 64]> value_layer_15_cast_fp16 = reshape(shape = var_1757, x = var_1753_cast_fp16)[name = string("value_layer_15_cast_fp16")];
             tensor<int32, [3]> var_1767_perm_0 = const()[name = string("op_1767_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1769_y_0_to_fp16 = const()[name = string("_inversed_1769_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = key_layer_15_cast_fp16)[name = string("transpose_98")];
-            tensor<fp16, [16, 64, 512]> _inversed_1769_cast_fp16 = mul(x = var_1767_cast_fp16, y = _inversed_1769_y_0_to_fp16)[name = string("_inversed_1769_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_1767_cast_fp16 = transpose(perm = var_1767_perm_0, x = key_layer_15_cast_fp16)[name = string("transpose_98")];
+            tensor<fp16, [16, 64, 224]> _inversed_1769_cast_fp16 = mul(x = var_1767_cast_fp16, y = _inversed_1769_y_0_to_fp16)[name = string("_inversed_1769_cast_fp16")];
             bool attention_scores_43_transpose_x_0 = const()[name = string("attention_scores_43_transpose_x_0"), val = bool(false)];
             bool attention_scores_43_transpose_y_0 = const()[name = string("attention_scores_43_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_43_cast_fp16 = matmul(transpose_x = attention_scores_43_transpose_x_0, transpose_y = attention_scores_43_transpose_y_0, x = query_layer_45_cast_fp16, y = _inversed_1769_cast_fp16)[name = string("attention_scores_43_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_43_cast_fp16 = matmul(transpose_x = attention_scores_43_transpose_x_0, transpose_y = attention_scores_43_transpose_y_0, x = query_layer_45_cast_fp16, y = _inversed_1769_cast_fp16)[name = string("attention_scores_43_cast_fp16")];
             bool c2p_att_29_transpose_x_0 = const()[name = string("c2p_att_29_transpose_x_0"), val = bool(false)];
             bool c2p_att_29_transpose_y_0 = const()[name = string("c2p_att_29_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1825_to_fp16 = const()[name = string("op_1825_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(460553408)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_29_cast_fp16 = matmul(transpose_x = c2p_att_29_transpose_x_0, transpose_y = c2p_att_29_transpose_y_0, x = query_layer_45_cast_fp16, y = var_1825_to_fp16)[name = string("c2p_att_29_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_29_cast_fp16 = matmul(transpose_x = c2p_att_29_transpose_x_0, transpose_y = c2p_att_29_transpose_y_0, x = query_layer_45_cast_fp16, y = var_1825_to_fp16)[name = string("c2p_att_29_cast_fp16")];
             bool c2p_att_31_validate_indices_0 = const()[name = string("c2p_att_31_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_31_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_31_validate_indices_0, x = c2p_att_29_cast_fp16)[name = string("c2p_att_31_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_31_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_31_validate_indices_0, x = c2p_att_29_cast_fp16)[name = string("c2p_att_31_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_1838_y_0_to_fp16 = const()[name = string("_inversed_1838_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_1838_cast_fp16 = mul(x = c2p_att_31_cast_fp16_cast_int16, y = _inversed_1838_y_0_to_fp16)[name = string("_inversed_1838_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_1838_cast_fp16 = mul(x = c2p_att_31_cast_fp16_cast_int16, y = _inversed_1838_y_0_to_fp16)[name = string("_inversed_1838_cast_fp16")];
             bool p2c_att_29_transpose_x_0 = const()[name = string("p2c_att_29_transpose_x_0"), val = bool(false)];
             bool p2c_att_29_transpose_y_0 = const()[name = string("p2c_att_29_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_1853_to_fp16 = const()[name = string("op_1853_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461602048)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_29_cast_fp16 = matmul(transpose_x = p2c_att_29_transpose_x_0, transpose_y = p2c_att_29_transpose_y_0, x = key_layer_15_cast_fp16, y = var_1853_to_fp16)[name = string("p2c_att_29_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_29_cast_fp16 = matmul(transpose_x = p2c_att_29_transpose_x_0, transpose_y = p2c_att_29_transpose_y_0, x = key_layer_15_cast_fp16, y = var_1853_to_fp16)[name = string("p2c_att_29_cast_fp16")];
             bool var_1861_validate_indices_0 = const()[name = string("op_1861_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_1861_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1861_validate_indices_0, x = p2c_att_29_cast_fp16)[name = string("op_1861_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_1861_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_1861_validate_indices_0, x = p2c_att_29_cast_fp16)[name = string("op_1861_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_31_perm_0 = const()[name = string("p2c_att_31_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1864_y_0_to_fp16 = const()[name = string("_inversed_1864_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_31_cast_fp16 = transpose(perm = p2c_att_31_perm_0, x = var_1861_cast_fp16_cast_int16)[name = string("transpose_97")];
-            tensor<fp16, [16, 512, 512]> _inversed_1864_cast_fp16 = mul(x = p2c_att_31_cast_fp16, y = _inversed_1864_y_0_to_fp16)[name = string("_inversed_1864_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_15_cast_fp16 = add(x = _inversed_1838_cast_fp16, y = _inversed_1864_cast_fp16)[name = string("rel_att_15_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_45_cast_fp16 = add(x = attention_scores_43_cast_fp16, y = rel_att_15_cast_fp16)[name = string("attention_scores_45_cast_fp16")];
-            tensor<int32, [4]> var_1869 = const()[name = string("op_1869"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_47_cast_fp16 = reshape(shape = var_1869, x = attention_scores_45_cast_fp16)[name = string("attention_scores_47_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_177_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_47_cast_fp16, cond = var_339)[name = string("input_177_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_179_cast_fp16 = softmax(axis = var_56, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
-            tensor<int32, [3]> var_1878 = const()[name = string("op_1878"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_1879_cast_fp16 = reshape(shape = var_1878, x = input_179_cast_fp16)[name = string("op_1879_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_31_cast_fp16 = transpose(perm = p2c_att_31_perm_0, x = var_1861_cast_fp16_cast_int16)[name = string("transpose_97")];
+            tensor<fp16, [16, 224, 224]> _inversed_1864_cast_fp16 = mul(x = p2c_att_31_cast_fp16, y = _inversed_1864_y_0_to_fp16)[name = string("_inversed_1864_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_15_cast_fp16 = add(x = _inversed_1838_cast_fp16, y = _inversed_1864_cast_fp16)[name = string("rel_att_15_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_45_cast_fp16 = add(x = attention_scores_43_cast_fp16, y = rel_att_15_cast_fp16)[name = string("attention_scores_45_cast_fp16")];
+            tensor<int32, [4]> var_1869 = const()[name = string("op_1869"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_47_cast_fp16 = reshape(shape = var_1869, x = attention_scores_45_cast_fp16)[name = string("attention_scores_47_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_177_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_47_cast_fp16, cond = var_339)[name = string("input_177_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_179_cast_fp16 = softmax(axis = var_56, x = input_177_cast_fp16)[name = string("input_179_cast_fp16")];
+            tensor<int32, [3]> var_1878 = const()[name = string("op_1878"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_1879_cast_fp16 = reshape(shape = var_1878, x = input_179_cast_fp16)[name = string("op_1879_cast_fp16")];
             bool context_layer_29_transpose_x_0 = const()[name = string("context_layer_29_transpose_x_0"), val = bool(false)];
             bool context_layer_29_transpose_y_0 = const()[name = string("context_layer_29_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = var_1879_cast_fp16, y = value_layer_15_cast_fp16)[name = string("context_layer_29_cast_fp16")];
-            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1884_cast_fp16 = reshape(shape = var_1883, x = context_layer_29_cast_fp16)[name = string("op_1884_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_29_cast_fp16 = matmul(transpose_x = context_layer_29_transpose_x_0, transpose_y = context_layer_29_transpose_y_0, x = var_1879_cast_fp16, y = value_layer_15_cast_fp16)[name = string("context_layer_29_cast_fp16")];
+            tensor<int32, [4]> var_1883 = const()[name = string("op_1883"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1884_cast_fp16 = reshape(shape = var_1883, x = context_layer_29_cast_fp16)[name = string("op_1884_cast_fp16")];
             tensor<int32, [4]> var_1885 = const()[name = string("op_1885"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1890 = const()[name = string("op_1890"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_1886_cast_fp16 = transpose(perm = var_1885, x = var_1884_cast_fp16)[name = string("transpose_96")];
-            tensor<fp16, [1, 512, 1024]> input_181_cast_fp16 = reshape(shape = var_1890, x = var_1886_cast_fp16)[name = string("input_181_cast_fp16")];
+            tensor<int32, [3]> var_1890 = const()[name = string("op_1890"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_1886_cast_fp16 = transpose(perm = var_1885, x = var_1884_cast_fp16)[name = string("transpose_96")];
+            tensor<fp16, [1, 224, 1024]> input_181_cast_fp16 = reshape(shape = var_1890, x = var_1886_cast_fp16)[name = string("input_181_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_7_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462650688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464747904)))];
-            tensor<fp16, [1, 512, 1024]> linear_61_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_61_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_185_cast_fp16 = add(x = linear_61_cast_fp16, y = input_173_cast_fp16)[name = string("input_185_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_61_cast_fp16 = linear(bias = encoder_encoder_layer_7_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_attention_output_dense_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_61_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_185_cast_fp16 = add(x = linear_61_cast_fp16, y = input_173_cast_fp16)[name = string("input_185_cast_fp16")];
             tensor<int32, [1]> input_187_axes_0 = const()[name = string("input_187_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464750016)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464752128)))];
-            tensor<fp16, [1, 512, 1024]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_187_cast_fp16 = layer_norm(axes = input_187_axes_0, beta = encoder_encoder_layer_7_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_attention_output_LayerNorm_weight_to_fp16, x = input_185_cast_fp16)[name = string("input_187_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_7_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(464754240)))];
             tensor<fp16, [4096]> encoder_encoder_layer_7_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473142912)))];
-            tensor<fp16, [1, 512, 4096]> linear_62_cast_fp16 = linear(bias = encoder_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_62_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_62_cast_fp16 = linear(bias = encoder_encoder_layer_7_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_7_intermediate_dense_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_62_cast_fp16")];
             string input_191_mode_0 = const()[name = string("input_191_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = string("input_191_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_191_cast_fp16 = gelu(mode = input_191_mode_0, x = linear_62_cast_fp16)[name = string("input_191_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_7_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473151168)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481539840)))];
-            tensor<fp16, [1, 512, 1024]> linear_63_cast_fp16 = linear(bias = encoder_encoder_layer_7_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_output_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_63_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_195_cast_fp16 = add(x = linear_63_cast_fp16, y = input_187_cast_fp16)[name = string("input_195_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_63_cast_fp16 = linear(bias = encoder_encoder_layer_7_output_dense_bias_to_fp16, weight = encoder_encoder_layer_7_output_dense_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_63_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_195_cast_fp16 = add(x = linear_63_cast_fp16, y = input_187_cast_fp16)[name = string("input_195_cast_fp16")];
             tensor<int32, [1]> input_197_axes_0 = const()[name = string("input_197_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481541952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481544064)))];
-            tensor<fp16, [1, 512, 1024]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_197_cast_fp16 = layer_norm(axes = input_197_axes_0, beta = encoder_encoder_layer_7_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_7_output_LayerNorm_weight_to_fp16, x = input_195_cast_fp16)[name = string("input_197_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481546176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483643392)))];
-            tensor<fp16, [1, 512, 1024]> linear_64_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_64_cast_fp16")];
-            tensor<int32, [4]> var_1941 = const()[name = string("op_1941"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_163_cast_fp16 = reshape(shape = var_1941, x = linear_64_cast_fp16)[name = string("x_163_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_64_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_query_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_64_cast_fp16")];
+            tensor<int32, [4]> var_1941 = const()[name = string("op_1941"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_163_cast_fp16 = reshape(shape = var_1941, x = linear_64_cast_fp16)[name = string("x_163_cast_fp16")];
             tensor<int32, [4]> var_1943 = const()[name = string("op_1943"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1948 = const()[name = string("op_1948"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1944_cast_fp16 = transpose(perm = var_1943, x = x_163_cast_fp16)[name = string("transpose_95")];
-            tensor<fp16, [16, 512, 64]> query_layer_51_cast_fp16 = reshape(shape = var_1948, x = var_1944_cast_fp16)[name = string("query_layer_51_cast_fp16")];
+            tensor<int32, [3]> var_1948 = const()[name = string("op_1948"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1944_cast_fp16 = transpose(perm = var_1943, x = x_163_cast_fp16)[name = string("transpose_95")];
+            tensor<fp16, [16, 224, 64]> query_layer_51_cast_fp16 = reshape(shape = var_1948, x = var_1944_cast_fp16)[name = string("query_layer_51_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(483645504)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485742720)))];
-            tensor<fp16, [1, 512, 1024]> linear_65_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_65_cast_fp16")];
-            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_167_cast_fp16 = reshape(shape = var_1955, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_65_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_key_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_65_cast_fp16")];
+            tensor<int32, [4]> var_1955 = const()[name = string("op_1955"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_167_cast_fp16 = reshape(shape = var_1955, x = linear_65_cast_fp16)[name = string("x_167_cast_fp16")];
             tensor<int32, [4]> var_1957 = const()[name = string("op_1957"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1962 = const()[name = string("op_1962"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1958_cast_fp16 = transpose(perm = var_1957, x = x_167_cast_fp16)[name = string("transpose_94")];
-            tensor<fp16, [16, 512, 64]> key_layer_17_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = string("key_layer_17_cast_fp16")];
+            tensor<int32, [3]> var_1962 = const()[name = string("op_1962"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1958_cast_fp16 = transpose(perm = var_1957, x = x_167_cast_fp16)[name = string("transpose_94")];
+            tensor<fp16, [16, 224, 64]> key_layer_17_cast_fp16 = reshape(shape = var_1962, x = var_1958_cast_fp16)[name = string("key_layer_17_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(485744832)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487842048)))];
-            tensor<fp16, [1, 512, 1024]> linear_66_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_66_cast_fp16")];
-            tensor<int32, [4]> var_1969 = const()[name = string("op_1969"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_171_cast_fp16 = reshape(shape = var_1969, x = linear_66_cast_fp16)[name = string("x_171_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_66_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_8_attention_self_value_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_66_cast_fp16")];
+            tensor<int32, [4]> var_1969 = const()[name = string("op_1969"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_171_cast_fp16 = reshape(shape = var_1969, x = linear_66_cast_fp16)[name = string("x_171_cast_fp16")];
             tensor<int32, [4]> var_1971 = const()[name = string("op_1971"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_1972_cast_fp16 = transpose(perm = var_1971, x = x_171_cast_fp16)[name = string("transpose_93")];
-            tensor<fp16, [16, 512, 64]> value_layer_17_cast_fp16 = reshape(shape = var_1976, x = var_1972_cast_fp16)[name = string("value_layer_17_cast_fp16")];
+            tensor<int32, [3]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_1972_cast_fp16 = transpose(perm = var_1971, x = x_171_cast_fp16)[name = string("transpose_93")];
+            tensor<fp16, [16, 224, 64]> value_layer_17_cast_fp16 = reshape(shape = var_1976, x = var_1972_cast_fp16)[name = string("value_layer_17_cast_fp16")];
             tensor<int32, [3]> var_1986_perm_0 = const()[name = string("op_1986_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_1988_y_0_to_fp16 = const()[name = string("_inversed_1988_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_1986_cast_fp16 = transpose(perm = var_1986_perm_0, x = key_layer_17_cast_fp16)[name = string("transpose_92")];
-            tensor<fp16, [16, 64, 512]> _inversed_1988_cast_fp16 = mul(x = var_1986_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = string("_inversed_1988_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_1986_cast_fp16 = transpose(perm = var_1986_perm_0, x = key_layer_17_cast_fp16)[name = string("transpose_92")];
+            tensor<fp16, [16, 64, 224]> _inversed_1988_cast_fp16 = mul(x = var_1986_cast_fp16, y = _inversed_1988_y_0_to_fp16)[name = string("_inversed_1988_cast_fp16")];
             bool attention_scores_49_transpose_x_0 = const()[name = string("attention_scores_49_transpose_x_0"), val = bool(false)];
             bool attention_scores_49_transpose_y_0 = const()[name = string("attention_scores_49_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_49_cast_fp16 = matmul(transpose_x = attention_scores_49_transpose_x_0, transpose_y = attention_scores_49_transpose_y_0, x = query_layer_51_cast_fp16, y = _inversed_1988_cast_fp16)[name = string("attention_scores_49_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_49_cast_fp16 = matmul(transpose_x = attention_scores_49_transpose_x_0, transpose_y = attention_scores_49_transpose_y_0, x = query_layer_51_cast_fp16, y = _inversed_1988_cast_fp16)[name = string("attention_scores_49_cast_fp16")];
             bool c2p_att_33_transpose_x_0 = const()[name = string("c2p_att_33_transpose_x_0"), val = bool(false)];
             bool c2p_att_33_transpose_y_0 = const()[name = string("c2p_att_33_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2044_to_fp16 = const()[name = string("op_2044_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487844160)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_33_cast_fp16 = matmul(transpose_x = c2p_att_33_transpose_x_0, transpose_y = c2p_att_33_transpose_y_0, x = query_layer_51_cast_fp16, y = var_2044_to_fp16)[name = string("c2p_att_33_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_33_cast_fp16 = matmul(transpose_x = c2p_att_33_transpose_x_0, transpose_y = c2p_att_33_transpose_y_0, x = query_layer_51_cast_fp16, y = var_2044_to_fp16)[name = string("c2p_att_33_cast_fp16")];
             bool c2p_att_35_validate_indices_0 = const()[name = string("c2p_att_35_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_35_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_35_validate_indices_0, x = c2p_att_33_cast_fp16)[name = string("c2p_att_35_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_35_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_35_validate_indices_0, x = c2p_att_33_cast_fp16)[name = string("c2p_att_35_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2057_y_0_to_fp16 = const()[name = string("_inversed_2057_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_2057_cast_fp16 = mul(x = c2p_att_35_cast_fp16_cast_int16, y = _inversed_2057_y_0_to_fp16)[name = string("_inversed_2057_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_2057_cast_fp16 = mul(x = c2p_att_35_cast_fp16_cast_int16, y = _inversed_2057_y_0_to_fp16)[name = string("_inversed_2057_cast_fp16")];
             bool p2c_att_33_transpose_x_0 = const()[name = string("p2c_att_33_transpose_x_0"), val = bool(false)];
             bool p2c_att_33_transpose_y_0 = const()[name = string("p2c_att_33_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2072_to_fp16 = const()[name = string("op_2072_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488892800)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_33_cast_fp16 = matmul(transpose_x = p2c_att_33_transpose_x_0, transpose_y = p2c_att_33_transpose_y_0, x = key_layer_17_cast_fp16, y = var_2072_to_fp16)[name = string("p2c_att_33_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_33_cast_fp16 = matmul(transpose_x = p2c_att_33_transpose_x_0, transpose_y = p2c_att_33_transpose_y_0, x = key_layer_17_cast_fp16, y = var_2072_to_fp16)[name = string("p2c_att_33_cast_fp16")];
             bool var_2080_validate_indices_0 = const()[name = string("op_2080_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_2080_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2080_validate_indices_0, x = p2c_att_33_cast_fp16)[name = string("op_2080_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_2080_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2080_validate_indices_0, x = p2c_att_33_cast_fp16)[name = string("op_2080_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_35_perm_0 = const()[name = string("p2c_att_35_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2083_y_0_to_fp16 = const()[name = string("_inversed_2083_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_35_cast_fp16 = transpose(perm = p2c_att_35_perm_0, x = var_2080_cast_fp16_cast_int16)[name = string("transpose_91")];
-            tensor<fp16, [16, 512, 512]> _inversed_2083_cast_fp16 = mul(x = p2c_att_35_cast_fp16, y = _inversed_2083_y_0_to_fp16)[name = string("_inversed_2083_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_17_cast_fp16 = add(x = _inversed_2057_cast_fp16, y = _inversed_2083_cast_fp16)[name = string("rel_att_17_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_51_cast_fp16 = add(x = attention_scores_49_cast_fp16, y = rel_att_17_cast_fp16)[name = string("attention_scores_51_cast_fp16")];
-            tensor<int32, [4]> var_2088 = const()[name = string("op_2088"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_53_cast_fp16 = reshape(shape = var_2088, x = attention_scores_51_cast_fp16)[name = string("attention_scores_53_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_201_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_53_cast_fp16, cond = var_339)[name = string("input_201_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_203_cast_fp16 = softmax(axis = var_56, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
-            tensor<int32, [3]> var_2097 = const()[name = string("op_2097"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_2098_cast_fp16 = reshape(shape = var_2097, x = input_203_cast_fp16)[name = string("op_2098_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_35_cast_fp16 = transpose(perm = p2c_att_35_perm_0, x = var_2080_cast_fp16_cast_int16)[name = string("transpose_91")];
+            tensor<fp16, [16, 224, 224]> _inversed_2083_cast_fp16 = mul(x = p2c_att_35_cast_fp16, y = _inversed_2083_y_0_to_fp16)[name = string("_inversed_2083_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_17_cast_fp16 = add(x = _inversed_2057_cast_fp16, y = _inversed_2083_cast_fp16)[name = string("rel_att_17_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_51_cast_fp16 = add(x = attention_scores_49_cast_fp16, y = rel_att_17_cast_fp16)[name = string("attention_scores_51_cast_fp16")];
+            tensor<int32, [4]> var_2088 = const()[name = string("op_2088"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_53_cast_fp16 = reshape(shape = var_2088, x = attention_scores_51_cast_fp16)[name = string("attention_scores_53_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_201_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_53_cast_fp16, cond = var_339)[name = string("input_201_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_203_cast_fp16 = softmax(axis = var_56, x = input_201_cast_fp16)[name = string("input_203_cast_fp16")];
+            tensor<int32, [3]> var_2097 = const()[name = string("op_2097"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_2098_cast_fp16 = reshape(shape = var_2097, x = input_203_cast_fp16)[name = string("op_2098_cast_fp16")];
             bool context_layer_33_transpose_x_0 = const()[name = string("context_layer_33_transpose_x_0"), val = bool(false)];
             bool context_layer_33_transpose_y_0 = const()[name = string("context_layer_33_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = var_2098_cast_fp16, y = value_layer_17_cast_fp16)[name = string("context_layer_33_cast_fp16")];
-            tensor<int32, [4]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2103_cast_fp16 = reshape(shape = var_2102, x = context_layer_33_cast_fp16)[name = string("op_2103_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_33_cast_fp16 = matmul(transpose_x = context_layer_33_transpose_x_0, transpose_y = context_layer_33_transpose_y_0, x = var_2098_cast_fp16, y = value_layer_17_cast_fp16)[name = string("context_layer_33_cast_fp16")];
+            tensor<int32, [4]> var_2102 = const()[name = string("op_2102"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2103_cast_fp16 = reshape(shape = var_2102, x = context_layer_33_cast_fp16)[name = string("op_2103_cast_fp16")];
             tensor<int32, [4]> var_2104 = const()[name = string("op_2104"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2109 = const()[name = string("op_2109"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_2105_cast_fp16 = transpose(perm = var_2104, x = var_2103_cast_fp16)[name = string("transpose_90")];
-            tensor<fp16, [1, 512, 1024]> input_205_cast_fp16 = reshape(shape = var_2109, x = var_2105_cast_fp16)[name = string("input_205_cast_fp16")];
+            tensor<int32, [3]> var_2109 = const()[name = string("op_2109"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_2105_cast_fp16 = transpose(perm = var_2104, x = var_2103_cast_fp16)[name = string("transpose_90")];
+            tensor<fp16, [1, 224, 1024]> input_205_cast_fp16 = reshape(shape = var_2109, x = var_2105_cast_fp16)[name = string("input_205_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_8_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(489941440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492038656)))];
-            tensor<fp16, [1, 512, 1024]> linear_69_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_69_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_197_cast_fp16)[name = string("input_209_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_69_cast_fp16 = linear(bias = encoder_encoder_layer_8_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_attention_output_dense_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_69_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_209_cast_fp16 = add(x = linear_69_cast_fp16, y = input_197_cast_fp16)[name = string("input_209_cast_fp16")];
             tensor<int32, [1]> input_211_axes_0 = const()[name = string("input_211_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492040768)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492042880)))];
-            tensor<fp16, [1, 512, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_211_cast_fp16 = layer_norm(axes = input_211_axes_0, beta = encoder_encoder_layer_8_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_attention_output_LayerNorm_weight_to_fp16, x = input_209_cast_fp16)[name = string("input_211_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_8_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(492044992)))];
             tensor<fp16, [4096]> encoder_encoder_layer_8_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500433664)))];
-            tensor<fp16, [1, 512, 4096]> linear_70_cast_fp16 = linear(bias = encoder_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_70_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_70_cast_fp16 = linear(bias = encoder_encoder_layer_8_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_8_intermediate_dense_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_70_cast_fp16")];
             string input_215_mode_0 = const()[name = string("input_215_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = string("input_215_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_215_cast_fp16 = gelu(mode = input_215_mode_0, x = linear_70_cast_fp16)[name = string("input_215_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_8_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500441920)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508830592)))];
-            tensor<fp16, [1, 512, 1024]> linear_71_cast_fp16 = linear(bias = encoder_encoder_layer_8_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_output_dense_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_71_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_211_cast_fp16)[name = string("input_219_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_71_cast_fp16 = linear(bias = encoder_encoder_layer_8_output_dense_bias_to_fp16, weight = encoder_encoder_layer_8_output_dense_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_71_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_219_cast_fp16 = add(x = linear_71_cast_fp16, y = input_211_cast_fp16)[name = string("input_219_cast_fp16")];
             tensor<int32, [1]> input_221_axes_0 = const()[name = string("input_221_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508832704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508834816)))];
-            tensor<fp16, [1, 512, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_221_cast_fp16 = layer_norm(axes = input_221_axes_0, beta = encoder_encoder_layer_8_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_8_output_LayerNorm_weight_to_fp16, x = input_219_cast_fp16)[name = string("input_221_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(508836928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510934144)))];
-            tensor<fp16, [1, 512, 1024]> linear_72_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_72_cast_fp16")];
-            tensor<int32, [4]> var_2160 = const()[name = string("op_2160"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_183_cast_fp16 = reshape(shape = var_2160, x = linear_72_cast_fp16)[name = string("x_183_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_72_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_query_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_72_cast_fp16")];
+            tensor<int32, [4]> var_2160 = const()[name = string("op_2160"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_183_cast_fp16 = reshape(shape = var_2160, x = linear_72_cast_fp16)[name = string("x_183_cast_fp16")];
             tensor<int32, [4]> var_2162 = const()[name = string("op_2162"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2167 = const()[name = string("op_2167"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2163_cast_fp16 = transpose(perm = var_2162, x = x_183_cast_fp16)[name = string("transpose_89")];
-            tensor<fp16, [16, 512, 64]> query_layer_57_cast_fp16 = reshape(shape = var_2167, x = var_2163_cast_fp16)[name = string("query_layer_57_cast_fp16")];
+            tensor<int32, [3]> var_2167 = const()[name = string("op_2167"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2163_cast_fp16 = transpose(perm = var_2162, x = x_183_cast_fp16)[name = string("transpose_89")];
+            tensor<fp16, [16, 224, 64]> query_layer_57_cast_fp16 = reshape(shape = var_2167, x = var_2163_cast_fp16)[name = string("query_layer_57_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(510936256)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513033472)))];
-            tensor<fp16, [1, 512, 1024]> linear_73_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_73_cast_fp16")];
-            tensor<int32, [4]> var_2174 = const()[name = string("op_2174"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_187_cast_fp16 = reshape(shape = var_2174, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_73_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_key_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_73_cast_fp16")];
+            tensor<int32, [4]> var_2174 = const()[name = string("op_2174"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_187_cast_fp16 = reshape(shape = var_2174, x = linear_73_cast_fp16)[name = string("x_187_cast_fp16")];
             tensor<int32, [4]> var_2176 = const()[name = string("op_2176"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2181 = const()[name = string("op_2181"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2177_cast_fp16 = transpose(perm = var_2176, x = x_187_cast_fp16)[name = string("transpose_88")];
-            tensor<fp16, [16, 512, 64]> key_layer_19_cast_fp16 = reshape(shape = var_2181, x = var_2177_cast_fp16)[name = string("key_layer_19_cast_fp16")];
+            tensor<int32, [3]> var_2181 = const()[name = string("op_2181"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2177_cast_fp16 = transpose(perm = var_2176, x = x_187_cast_fp16)[name = string("transpose_88")];
+            tensor<fp16, [16, 224, 64]> key_layer_19_cast_fp16 = reshape(shape = var_2181, x = var_2177_cast_fp16)[name = string("key_layer_19_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(513035584)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515132800)))];
-            tensor<fp16, [1, 512, 1024]> linear_74_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_74_cast_fp16")];
-            tensor<int32, [4]> var_2188 = const()[name = string("op_2188"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_191_cast_fp16 = reshape(shape = var_2188, x = linear_74_cast_fp16)[name = string("x_191_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_74_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_9_attention_self_value_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_74_cast_fp16")];
+            tensor<int32, [4]> var_2188 = const()[name = string("op_2188"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_191_cast_fp16 = reshape(shape = var_2188, x = linear_74_cast_fp16)[name = string("x_191_cast_fp16")];
             tensor<int32, [4]> var_2190 = const()[name = string("op_2190"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2195 = const()[name = string("op_2195"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2191_cast_fp16 = transpose(perm = var_2190, x = x_191_cast_fp16)[name = string("transpose_87")];
-            tensor<fp16, [16, 512, 64]> value_layer_19_cast_fp16 = reshape(shape = var_2195, x = var_2191_cast_fp16)[name = string("value_layer_19_cast_fp16")];
+            tensor<int32, [3]> var_2195 = const()[name = string("op_2195"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2191_cast_fp16 = transpose(perm = var_2190, x = x_191_cast_fp16)[name = string("transpose_87")];
+            tensor<fp16, [16, 224, 64]> value_layer_19_cast_fp16 = reshape(shape = var_2195, x = var_2191_cast_fp16)[name = string("value_layer_19_cast_fp16")];
             tensor<int32, [3]> var_2205_perm_0 = const()[name = string("op_2205_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2207_y_0_to_fp16 = const()[name = string("_inversed_2207_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = key_layer_19_cast_fp16)[name = string("transpose_86")];
-            tensor<fp16, [16, 64, 512]> _inversed_2207_cast_fp16 = mul(x = var_2205_cast_fp16, y = _inversed_2207_y_0_to_fp16)[name = string("_inversed_2207_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_2205_cast_fp16 = transpose(perm = var_2205_perm_0, x = key_layer_19_cast_fp16)[name = string("transpose_86")];
+            tensor<fp16, [16, 64, 224]> _inversed_2207_cast_fp16 = mul(x = var_2205_cast_fp16, y = _inversed_2207_y_0_to_fp16)[name = string("_inversed_2207_cast_fp16")];
             bool attention_scores_55_transpose_x_0 = const()[name = string("attention_scores_55_transpose_x_0"), val = bool(false)];
             bool attention_scores_55_transpose_y_0 = const()[name = string("attention_scores_55_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_55_cast_fp16 = matmul(transpose_x = attention_scores_55_transpose_x_0, transpose_y = attention_scores_55_transpose_y_0, x = query_layer_57_cast_fp16, y = _inversed_2207_cast_fp16)[name = string("attention_scores_55_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_55_cast_fp16 = matmul(transpose_x = attention_scores_55_transpose_x_0, transpose_y = attention_scores_55_transpose_y_0, x = query_layer_57_cast_fp16, y = _inversed_2207_cast_fp16)[name = string("attention_scores_55_cast_fp16")];
             bool c2p_att_37_transpose_x_0 = const()[name = string("c2p_att_37_transpose_x_0"), val = bool(false)];
             bool c2p_att_37_transpose_y_0 = const()[name = string("c2p_att_37_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2263_to_fp16 = const()[name = string("op_2263_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(515134912)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_37_cast_fp16 = matmul(transpose_x = c2p_att_37_transpose_x_0, transpose_y = c2p_att_37_transpose_y_0, x = query_layer_57_cast_fp16, y = var_2263_to_fp16)[name = string("c2p_att_37_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_37_cast_fp16 = matmul(transpose_x = c2p_att_37_transpose_x_0, transpose_y = c2p_att_37_transpose_y_0, x = query_layer_57_cast_fp16, y = var_2263_to_fp16)[name = string("c2p_att_37_cast_fp16")];
             bool c2p_att_39_validate_indices_0 = const()[name = string("c2p_att_39_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_39_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_39_validate_indices_0, x = c2p_att_37_cast_fp16)[name = string("c2p_att_39_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_39_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_39_validate_indices_0, x = c2p_att_37_cast_fp16)[name = string("c2p_att_39_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2276_y_0_to_fp16 = const()[name = string("_inversed_2276_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_2276_cast_fp16 = mul(x = c2p_att_39_cast_fp16_cast_int16, y = _inversed_2276_y_0_to_fp16)[name = string("_inversed_2276_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_2276_cast_fp16 = mul(x = c2p_att_39_cast_fp16_cast_int16, y = _inversed_2276_y_0_to_fp16)[name = string("_inversed_2276_cast_fp16")];
             bool p2c_att_37_transpose_x_0 = const()[name = string("p2c_att_37_transpose_x_0"), val = bool(false)];
             bool p2c_att_37_transpose_y_0 = const()[name = string("p2c_att_37_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2291_to_fp16 = const()[name = string("op_2291_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516183552)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_37_cast_fp16 = matmul(transpose_x = p2c_att_37_transpose_x_0, transpose_y = p2c_att_37_transpose_y_0, x = key_layer_19_cast_fp16, y = var_2291_to_fp16)[name = string("p2c_att_37_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_37_cast_fp16 = matmul(transpose_x = p2c_att_37_transpose_x_0, transpose_y = p2c_att_37_transpose_y_0, x = key_layer_19_cast_fp16, y = var_2291_to_fp16)[name = string("p2c_att_37_cast_fp16")];
             bool var_2299_validate_indices_0 = const()[name = string("op_2299_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_2299_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2299_validate_indices_0, x = p2c_att_37_cast_fp16)[name = string("op_2299_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_2299_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2299_validate_indices_0, x = p2c_att_37_cast_fp16)[name = string("op_2299_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_39_perm_0 = const()[name = string("p2c_att_39_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2302_y_0_to_fp16 = const()[name = string("_inversed_2302_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_39_cast_fp16 = transpose(perm = p2c_att_39_perm_0, x = var_2299_cast_fp16_cast_int16)[name = string("transpose_85")];
-            tensor<fp16, [16, 512, 512]> _inversed_2302_cast_fp16 = mul(x = p2c_att_39_cast_fp16, y = _inversed_2302_y_0_to_fp16)[name = string("_inversed_2302_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_19_cast_fp16 = add(x = _inversed_2276_cast_fp16, y = _inversed_2302_cast_fp16)[name = string("rel_att_19_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_57_cast_fp16 = add(x = attention_scores_55_cast_fp16, y = rel_att_19_cast_fp16)[name = string("attention_scores_57_cast_fp16")];
-            tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_59_cast_fp16 = reshape(shape = var_2307, x = attention_scores_57_cast_fp16)[name = string("attention_scores_59_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_225_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_59_cast_fp16, cond = var_339)[name = string("input_225_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_227_cast_fp16 = softmax(axis = var_56, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
-            tensor<int32, [3]> var_2316 = const()[name = string("op_2316"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_2317_cast_fp16 = reshape(shape = var_2316, x = input_227_cast_fp16)[name = string("op_2317_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_39_cast_fp16 = transpose(perm = p2c_att_39_perm_0, x = var_2299_cast_fp16_cast_int16)[name = string("transpose_85")];
+            tensor<fp16, [16, 224, 224]> _inversed_2302_cast_fp16 = mul(x = p2c_att_39_cast_fp16, y = _inversed_2302_y_0_to_fp16)[name = string("_inversed_2302_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_19_cast_fp16 = add(x = _inversed_2276_cast_fp16, y = _inversed_2302_cast_fp16)[name = string("rel_att_19_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_57_cast_fp16 = add(x = attention_scores_55_cast_fp16, y = rel_att_19_cast_fp16)[name = string("attention_scores_57_cast_fp16")];
+            tensor<int32, [4]> var_2307 = const()[name = string("op_2307"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_59_cast_fp16 = reshape(shape = var_2307, x = attention_scores_57_cast_fp16)[name = string("attention_scores_59_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_225_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_59_cast_fp16, cond = var_339)[name = string("input_225_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_227_cast_fp16 = softmax(axis = var_56, x = input_225_cast_fp16)[name = string("input_227_cast_fp16")];
+            tensor<int32, [3]> var_2316 = const()[name = string("op_2316"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_2317_cast_fp16 = reshape(shape = var_2316, x = input_227_cast_fp16)[name = string("op_2317_cast_fp16")];
             bool context_layer_37_transpose_x_0 = const()[name = string("context_layer_37_transpose_x_0"), val = bool(false)];
             bool context_layer_37_transpose_y_0 = const()[name = string("context_layer_37_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = var_2317_cast_fp16, y = value_layer_19_cast_fp16)[name = string("context_layer_37_cast_fp16")];
-            tensor<int32, [4]> var_2321 = const()[name = string("op_2321"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2322_cast_fp16 = reshape(shape = var_2321, x = context_layer_37_cast_fp16)[name = string("op_2322_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_37_cast_fp16 = matmul(transpose_x = context_layer_37_transpose_x_0, transpose_y = context_layer_37_transpose_y_0, x = var_2317_cast_fp16, y = value_layer_19_cast_fp16)[name = string("context_layer_37_cast_fp16")];
+            tensor<int32, [4]> var_2321 = const()[name = string("op_2321"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2322_cast_fp16 = reshape(shape = var_2321, x = context_layer_37_cast_fp16)[name = string("op_2322_cast_fp16")];
             tensor<int32, [4]> var_2323 = const()[name = string("op_2323"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_2324_cast_fp16 = transpose(perm = var_2323, x = var_2322_cast_fp16)[name = string("transpose_84")];
-            tensor<fp16, [1, 512, 1024]> input_229_cast_fp16 = reshape(shape = var_2328, x = var_2324_cast_fp16)[name = string("input_229_cast_fp16")];
+            tensor<int32, [3]> var_2328 = const()[name = string("op_2328"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_2324_cast_fp16 = transpose(perm = var_2323, x = var_2322_cast_fp16)[name = string("transpose_84")];
+            tensor<fp16, [1, 224, 1024]> input_229_cast_fp16 = reshape(shape = var_2328, x = var_2324_cast_fp16)[name = string("input_229_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_9_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(517232192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519329408)))];
-            tensor<fp16, [1, 512, 1024]> linear_77_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_77_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_233_cast_fp16 = add(x = linear_77_cast_fp16, y = input_221_cast_fp16)[name = string("input_233_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_77_cast_fp16 = linear(bias = encoder_encoder_layer_9_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_attention_output_dense_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_77_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_233_cast_fp16 = add(x = linear_77_cast_fp16, y = input_221_cast_fp16)[name = string("input_233_cast_fp16")];
             tensor<int32, [1]> input_235_axes_0 = const()[name = string("input_235_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519331520)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519333632)))];
-            tensor<fp16, [1, 512, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_235_cast_fp16 = layer_norm(axes = input_235_axes_0, beta = encoder_encoder_layer_9_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_attention_output_LayerNorm_weight_to_fp16, x = input_233_cast_fp16)[name = string("input_235_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_9_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519335744)))];
             tensor<fp16, [4096]> encoder_encoder_layer_9_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527724416)))];
-            tensor<fp16, [1, 512, 4096]> linear_78_cast_fp16 = linear(bias = encoder_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_78_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_78_cast_fp16 = linear(bias = encoder_encoder_layer_9_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_9_intermediate_dense_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_78_cast_fp16")];
             string input_239_mode_0 = const()[name = string("input_239_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = string("input_239_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_239_cast_fp16 = gelu(mode = input_239_mode_0, x = linear_78_cast_fp16)[name = string("input_239_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_9_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(527732672)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536121344)))];
-            tensor<fp16, [1, 512, 1024]> linear_79_cast_fp16 = linear(bias = encoder_encoder_layer_9_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_output_dense_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_79_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_243_cast_fp16 = add(x = linear_79_cast_fp16, y = input_235_cast_fp16)[name = string("input_243_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_79_cast_fp16 = linear(bias = encoder_encoder_layer_9_output_dense_bias_to_fp16, weight = encoder_encoder_layer_9_output_dense_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_79_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_243_cast_fp16 = add(x = linear_79_cast_fp16, y = input_235_cast_fp16)[name = string("input_243_cast_fp16")];
             tensor<int32, [1]> input_245_axes_0 = const()[name = string("input_245_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536123456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536125568)))];
-            tensor<fp16, [1, 512, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_245_cast_fp16 = layer_norm(axes = input_245_axes_0, beta = encoder_encoder_layer_9_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_9_output_LayerNorm_weight_to_fp16, x = input_243_cast_fp16)[name = string("input_245_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536127680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538224896)))];
-            tensor<fp16, [1, 512, 1024]> linear_80_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_80_cast_fp16")];
-            tensor<int32, [4]> var_2379 = const()[name = string("op_2379"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_203_cast_fp16 = reshape(shape = var_2379, x = linear_80_cast_fp16)[name = string("x_203_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_80_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_query_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_80_cast_fp16")];
+            tensor<int32, [4]> var_2379 = const()[name = string("op_2379"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_203_cast_fp16 = reshape(shape = var_2379, x = linear_80_cast_fp16)[name = string("x_203_cast_fp16")];
             tensor<int32, [4]> var_2381 = const()[name = string("op_2381"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2382_cast_fp16 = transpose(perm = var_2381, x = x_203_cast_fp16)[name = string("transpose_83")];
-            tensor<fp16, [16, 512, 64]> query_layer_63_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("query_layer_63_cast_fp16")];
+            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2382_cast_fp16 = transpose(perm = var_2381, x = x_203_cast_fp16)[name = string("transpose_83")];
+            tensor<fp16, [16, 224, 64]> query_layer_63_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("query_layer_63_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538227008)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540324224)))];
-            tensor<fp16, [1, 512, 1024]> linear_81_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_81_cast_fp16")];
-            tensor<int32, [4]> var_2393 = const()[name = string("op_2393"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_207_cast_fp16 = reshape(shape = var_2393, x = linear_81_cast_fp16)[name = string("x_207_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_81_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_key_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_81_cast_fp16")];
+            tensor<int32, [4]> var_2393 = const()[name = string("op_2393"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_207_cast_fp16 = reshape(shape = var_2393, x = linear_81_cast_fp16)[name = string("x_207_cast_fp16")];
             tensor<int32, [4]> var_2395 = const()[name = string("op_2395"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2400 = const()[name = string("op_2400"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2396_cast_fp16 = transpose(perm = var_2395, x = x_207_cast_fp16)[name = string("transpose_82")];
-            tensor<fp16, [16, 512, 64]> key_layer_21_cast_fp16 = reshape(shape = var_2400, x = var_2396_cast_fp16)[name = string("key_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_2400 = const()[name = string("op_2400"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2396_cast_fp16 = transpose(perm = var_2395, x = x_207_cast_fp16)[name = string("transpose_82")];
+            tensor<fp16, [16, 224, 64]> key_layer_21_cast_fp16 = reshape(shape = var_2400, x = var_2396_cast_fp16)[name = string("key_layer_21_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(540326336)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542423552)))];
-            tensor<fp16, [1, 512, 1024]> linear_82_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_82_cast_fp16")];
-            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_211_cast_fp16 = reshape(shape = var_2407, x = linear_82_cast_fp16)[name = string("x_211_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_82_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_10_attention_self_value_proj_weight_to_fp16, x = input_245_cast_fp16)[name = string("linear_82_cast_fp16")];
+            tensor<int32, [4]> var_2407 = const()[name = string("op_2407"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_211_cast_fp16 = reshape(shape = var_2407, x = linear_82_cast_fp16)[name = string("x_211_cast_fp16")];
             tensor<int32, [4]> var_2409 = const()[name = string("op_2409"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2414 = const()[name = string("op_2414"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2410_cast_fp16 = transpose(perm = var_2409, x = x_211_cast_fp16)[name = string("transpose_81")];
-            tensor<fp16, [16, 512, 64]> value_layer_21_cast_fp16 = reshape(shape = var_2414, x = var_2410_cast_fp16)[name = string("value_layer_21_cast_fp16")];
+            tensor<int32, [3]> var_2414 = const()[name = string("op_2414"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2410_cast_fp16 = transpose(perm = var_2409, x = x_211_cast_fp16)[name = string("transpose_81")];
+            tensor<fp16, [16, 224, 64]> value_layer_21_cast_fp16 = reshape(shape = var_2414, x = var_2410_cast_fp16)[name = string("value_layer_21_cast_fp16")];
             tensor<int32, [3]> var_2424_perm_0 = const()[name = string("op_2424_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2426_y_0_to_fp16 = const()[name = string("_inversed_2426_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_2424_cast_fp16 = transpose(perm = var_2424_perm_0, x = key_layer_21_cast_fp16)[name = string("transpose_80")];
-            tensor<fp16, [16, 64, 512]> _inversed_2426_cast_fp16 = mul(x = var_2424_cast_fp16, y = _inversed_2426_y_0_to_fp16)[name = string("_inversed_2426_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_2424_cast_fp16 = transpose(perm = var_2424_perm_0, x = key_layer_21_cast_fp16)[name = string("transpose_80")];
+            tensor<fp16, [16, 64, 224]> _inversed_2426_cast_fp16 = mul(x = var_2424_cast_fp16, y = _inversed_2426_y_0_to_fp16)[name = string("_inversed_2426_cast_fp16")];
             bool attention_scores_61_transpose_x_0 = const()[name = string("attention_scores_61_transpose_x_0"), val = bool(false)];
             bool attention_scores_61_transpose_y_0 = const()[name = string("attention_scores_61_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_61_cast_fp16 = matmul(transpose_x = attention_scores_61_transpose_x_0, transpose_y = attention_scores_61_transpose_y_0, x = query_layer_63_cast_fp16, y = _inversed_2426_cast_fp16)[name = string("attention_scores_61_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_61_cast_fp16 = matmul(transpose_x = attention_scores_61_transpose_x_0, transpose_y = attention_scores_61_transpose_y_0, x = query_layer_63_cast_fp16, y = _inversed_2426_cast_fp16)[name = string("attention_scores_61_cast_fp16")];
             bool c2p_att_41_transpose_x_0 = const()[name = string("c2p_att_41_transpose_x_0"), val = bool(false)];
             bool c2p_att_41_transpose_y_0 = const()[name = string("c2p_att_41_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2482_to_fp16 = const()[name = string("op_2482_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(542425664)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_41_cast_fp16 = matmul(transpose_x = c2p_att_41_transpose_x_0, transpose_y = c2p_att_41_transpose_y_0, x = query_layer_63_cast_fp16, y = var_2482_to_fp16)[name = string("c2p_att_41_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_41_cast_fp16 = matmul(transpose_x = c2p_att_41_transpose_x_0, transpose_y = c2p_att_41_transpose_y_0, x = query_layer_63_cast_fp16, y = var_2482_to_fp16)[name = string("c2p_att_41_cast_fp16")];
             bool c2p_att_43_validate_indices_0 = const()[name = string("c2p_att_43_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_43_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_43_validate_indices_0, x = c2p_att_41_cast_fp16)[name = string("c2p_att_43_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_43_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_43_validate_indices_0, x = c2p_att_41_cast_fp16)[name = string("c2p_att_43_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2495_y_0_to_fp16 = const()[name = string("_inversed_2495_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_2495_cast_fp16 = mul(x = c2p_att_43_cast_fp16_cast_int16, y = _inversed_2495_y_0_to_fp16)[name = string("_inversed_2495_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_2495_cast_fp16 = mul(x = c2p_att_43_cast_fp16_cast_int16, y = _inversed_2495_y_0_to_fp16)[name = string("_inversed_2495_cast_fp16")];
             bool p2c_att_41_transpose_x_0 = const()[name = string("p2c_att_41_transpose_x_0"), val = bool(false)];
             bool p2c_att_41_transpose_y_0 = const()[name = string("p2c_att_41_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2510_to_fp16 = const()[name = string("op_2510_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(543474304)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_41_cast_fp16 = matmul(transpose_x = p2c_att_41_transpose_x_0, transpose_y = p2c_att_41_transpose_y_0, x = key_layer_21_cast_fp16, y = var_2510_to_fp16)[name = string("p2c_att_41_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_41_cast_fp16 = matmul(transpose_x = p2c_att_41_transpose_x_0, transpose_y = p2c_att_41_transpose_y_0, x = key_layer_21_cast_fp16, y = var_2510_to_fp16)[name = string("p2c_att_41_cast_fp16")];
             bool var_2518_validate_indices_0 = const()[name = string("op_2518_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_2518_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2518_validate_indices_0, x = p2c_att_41_cast_fp16)[name = string("op_2518_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_2518_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2518_validate_indices_0, x = p2c_att_41_cast_fp16)[name = string("op_2518_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_43_perm_0 = const()[name = string("p2c_att_43_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2521_y_0_to_fp16 = const()[name = string("_inversed_2521_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_43_cast_fp16 = transpose(perm = p2c_att_43_perm_0, x = var_2518_cast_fp16_cast_int16)[name = string("transpose_79")];
-            tensor<fp16, [16, 512, 512]> _inversed_2521_cast_fp16 = mul(x = p2c_att_43_cast_fp16, y = _inversed_2521_y_0_to_fp16)[name = string("_inversed_2521_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_21_cast_fp16 = add(x = _inversed_2495_cast_fp16, y = _inversed_2521_cast_fp16)[name = string("rel_att_21_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_63_cast_fp16 = add(x = attention_scores_61_cast_fp16, y = rel_att_21_cast_fp16)[name = string("attention_scores_63_cast_fp16")];
-            tensor<int32, [4]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_65_cast_fp16 = reshape(shape = var_2526, x = attention_scores_63_cast_fp16)[name = string("attention_scores_65_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_249_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_65_cast_fp16, cond = var_339)[name = string("input_249_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_251_cast_fp16 = softmax(axis = var_56, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
-            tensor<int32, [3]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_2536_cast_fp16 = reshape(shape = var_2535, x = input_251_cast_fp16)[name = string("op_2536_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_43_cast_fp16 = transpose(perm = p2c_att_43_perm_0, x = var_2518_cast_fp16_cast_int16)[name = string("transpose_79")];
+            tensor<fp16, [16, 224, 224]> _inversed_2521_cast_fp16 = mul(x = p2c_att_43_cast_fp16, y = _inversed_2521_y_0_to_fp16)[name = string("_inversed_2521_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_21_cast_fp16 = add(x = _inversed_2495_cast_fp16, y = _inversed_2521_cast_fp16)[name = string("rel_att_21_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_63_cast_fp16 = add(x = attention_scores_61_cast_fp16, y = rel_att_21_cast_fp16)[name = string("attention_scores_63_cast_fp16")];
+            tensor<int32, [4]> var_2526 = const()[name = string("op_2526"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_65_cast_fp16 = reshape(shape = var_2526, x = attention_scores_63_cast_fp16)[name = string("attention_scores_65_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_249_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_65_cast_fp16, cond = var_339)[name = string("input_249_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_251_cast_fp16 = softmax(axis = var_56, x = input_249_cast_fp16)[name = string("input_251_cast_fp16")];
+            tensor<int32, [3]> var_2535 = const()[name = string("op_2535"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_2536_cast_fp16 = reshape(shape = var_2535, x = input_251_cast_fp16)[name = string("op_2536_cast_fp16")];
             bool context_layer_41_transpose_x_0 = const()[name = string("context_layer_41_transpose_x_0"), val = bool(false)];
             bool context_layer_41_transpose_y_0 = const()[name = string("context_layer_41_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = var_2536_cast_fp16, y = value_layer_21_cast_fp16)[name = string("context_layer_41_cast_fp16")];
-            tensor<int32, [4]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2541_cast_fp16 = reshape(shape = var_2540, x = context_layer_41_cast_fp16)[name = string("op_2541_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_41_cast_fp16 = matmul(transpose_x = context_layer_41_transpose_x_0, transpose_y = context_layer_41_transpose_y_0, x = var_2536_cast_fp16, y = value_layer_21_cast_fp16)[name = string("context_layer_41_cast_fp16")];
+            tensor<int32, [4]> var_2540 = const()[name = string("op_2540"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2541_cast_fp16 = reshape(shape = var_2540, x = context_layer_41_cast_fp16)[name = string("op_2541_cast_fp16")];
             tensor<int32, [4]> var_2542 = const()[name = string("op_2542"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_2543_cast_fp16 = transpose(perm = var_2542, x = var_2541_cast_fp16)[name = string("transpose_78")];
-            tensor<fp16, [1, 512, 1024]> input_253_cast_fp16 = reshape(shape = var_2547, x = var_2543_cast_fp16)[name = string("input_253_cast_fp16")];
+            tensor<int32, [3]> var_2547 = const()[name = string("op_2547"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_2543_cast_fp16 = transpose(perm = var_2542, x = var_2541_cast_fp16)[name = string("transpose_78")];
+            tensor<fp16, [1, 224, 1024]> input_253_cast_fp16 = reshape(shape = var_2547, x = var_2543_cast_fp16)[name = string("input_253_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_10_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544522944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546620160)))];
-            tensor<fp16, [1, 512, 1024]> linear_85_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_253_cast_fp16)[name = string("linear_85_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_257_cast_fp16 = add(x = linear_85_cast_fp16, y = input_245_cast_fp16)[name = string("input_257_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_85_cast_fp16 = linear(bias = encoder_encoder_layer_10_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_attention_output_dense_weight_to_fp16, x = input_253_cast_fp16)[name = string("linear_85_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_257_cast_fp16 = add(x = linear_85_cast_fp16, y = input_245_cast_fp16)[name = string("input_257_cast_fp16")];
             tensor<int32, [1]> input_259_axes_0 = const()[name = string("input_259_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546622272)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546624384)))];
-            tensor<fp16, [1, 512, 1024]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_259_cast_fp16 = layer_norm(axes = input_259_axes_0, beta = encoder_encoder_layer_10_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_attention_output_LayerNorm_weight_to_fp16, x = input_257_cast_fp16)[name = string("input_259_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_10_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(546626496)))];
             tensor<fp16, [4096]> encoder_encoder_layer_10_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555015168)))];
-            tensor<fp16, [1, 512, 4096]> linear_86_cast_fp16 = linear(bias = encoder_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_86_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_86_cast_fp16 = linear(bias = encoder_encoder_layer_10_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_10_intermediate_dense_weight_to_fp16, x = input_259_cast_fp16)[name = string("linear_86_cast_fp16")];
             string input_263_mode_0 = const()[name = string("input_263_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = string("input_263_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_263_cast_fp16 = gelu(mode = input_263_mode_0, x = linear_86_cast_fp16)[name = string("input_263_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_10_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(555023424)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563412096)))];
-            tensor<fp16, [1, 512, 1024]> linear_87_cast_fp16 = linear(bias = encoder_encoder_layer_10_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_output_dense_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_87_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_267_cast_fp16 = add(x = linear_87_cast_fp16, y = input_259_cast_fp16)[name = string("input_267_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_87_cast_fp16 = linear(bias = encoder_encoder_layer_10_output_dense_bias_to_fp16, weight = encoder_encoder_layer_10_output_dense_weight_to_fp16, x = input_263_cast_fp16)[name = string("linear_87_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_267_cast_fp16 = add(x = linear_87_cast_fp16, y = input_259_cast_fp16)[name = string("input_267_cast_fp16")];
             tensor<int32, [1]> input_269_axes_0 = const()[name = string("input_269_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563414208)))];
             tensor<fp16, [1024]> encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563416320)))];
-            tensor<fp16, [1, 512, 1024]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_269_cast_fp16 = layer_norm(axes = input_269_axes_0, beta = encoder_encoder_layer_10_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_10_output_LayerNorm_weight_to_fp16, x = input_267_cast_fp16)[name = string("input_269_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563418432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565515648)))];
-            tensor<fp16, [1, 512, 1024]> linear_88_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_88_cast_fp16")];
-            tensor<int32, [4]> var_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_223_cast_fp16 = reshape(shape = var_2598, x = linear_88_cast_fp16)[name = string("x_223_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_88_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_query_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_88_cast_fp16")];
+            tensor<int32, [4]> var_2598 = const()[name = string("op_2598"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_223_cast_fp16 = reshape(shape = var_2598, x = linear_88_cast_fp16)[name = string("x_223_cast_fp16")];
             tensor<int32, [4]> var_2600 = const()[name = string("op_2600"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2601_cast_fp16 = transpose(perm = var_2600, x = x_223_cast_fp16)[name = string("transpose_77")];
-            tensor<fp16, [16, 512, 64]> query_layer_69_cast_fp16 = reshape(shape = var_2605, x = var_2601_cast_fp16)[name = string("query_layer_69_cast_fp16")];
+            tensor<int32, [3]> var_2605 = const()[name = string("op_2605"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2601_cast_fp16 = transpose(perm = var_2600, x = x_223_cast_fp16)[name = string("transpose_77")];
+            tensor<fp16, [16, 224, 64]> query_layer_69_cast_fp16 = reshape(shape = var_2605, x = var_2601_cast_fp16)[name = string("query_layer_69_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565517760)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567614976)))];
-            tensor<fp16, [1, 512, 1024]> linear_89_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_89_cast_fp16")];
-            tensor<int32, [4]> var_2612 = const()[name = string("op_2612"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_227_cast_fp16 = reshape(shape = var_2612, x = linear_89_cast_fp16)[name = string("x_227_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_89_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_key_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_89_cast_fp16")];
+            tensor<int32, [4]> var_2612 = const()[name = string("op_2612"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_227_cast_fp16 = reshape(shape = var_2612, x = linear_89_cast_fp16)[name = string("x_227_cast_fp16")];
             tensor<int32, [4]> var_2614 = const()[name = string("op_2614"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2615_cast_fp16 = transpose(perm = var_2614, x = x_227_cast_fp16)[name = string("transpose_76")];
-            tensor<fp16, [16, 512, 64]> key_layer_23_cast_fp16 = reshape(shape = var_2619, x = var_2615_cast_fp16)[name = string("key_layer_23_cast_fp16")];
+            tensor<int32, [3]> var_2619 = const()[name = string("op_2619"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2615_cast_fp16 = transpose(perm = var_2614, x = x_227_cast_fp16)[name = string("transpose_76")];
+            tensor<fp16, [16, 224, 64]> key_layer_23_cast_fp16 = reshape(shape = var_2619, x = var_2615_cast_fp16)[name = string("key_layer_23_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567617088)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569714304)))];
-            tensor<fp16, [1, 512, 1024]> linear_90_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_90_cast_fp16")];
-            tensor<int32, [4]> var_2626 = const()[name = string("op_2626"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_231_cast_fp16 = reshape(shape = var_2626, x = linear_90_cast_fp16)[name = string("x_231_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_90_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_11_attention_self_value_proj_weight_to_fp16, x = input_269_cast_fp16)[name = string("linear_90_cast_fp16")];
+            tensor<int32, [4]> var_2626 = const()[name = string("op_2626"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_231_cast_fp16 = reshape(shape = var_2626, x = linear_90_cast_fp16)[name = string("x_231_cast_fp16")];
             tensor<int32, [4]> var_2628 = const()[name = string("op_2628"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2633 = const()[name = string("op_2633"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2629_cast_fp16 = transpose(perm = var_2628, x = x_231_cast_fp16)[name = string("transpose_75")];
-            tensor<fp16, [16, 512, 64]> value_layer_23_cast_fp16 = reshape(shape = var_2633, x = var_2629_cast_fp16)[name = string("value_layer_23_cast_fp16")];
+            tensor<int32, [3]> var_2633 = const()[name = string("op_2633"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2629_cast_fp16 = transpose(perm = var_2628, x = x_231_cast_fp16)[name = string("transpose_75")];
+            tensor<fp16, [16, 224, 64]> value_layer_23_cast_fp16 = reshape(shape = var_2633, x = var_2629_cast_fp16)[name = string("value_layer_23_cast_fp16")];
             tensor<int32, [3]> var_2643_perm_0 = const()[name = string("op_2643_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2645_y_0_to_fp16 = const()[name = string("_inversed_2645_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_2643_cast_fp16 = transpose(perm = var_2643_perm_0, x = key_layer_23_cast_fp16)[name = string("transpose_74")];
-            tensor<fp16, [16, 64, 512]> _inversed_2645_cast_fp16 = mul(x = var_2643_cast_fp16, y = _inversed_2645_y_0_to_fp16)[name = string("_inversed_2645_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_2643_cast_fp16 = transpose(perm = var_2643_perm_0, x = key_layer_23_cast_fp16)[name = string("transpose_74")];
+            tensor<fp16, [16, 64, 224]> _inversed_2645_cast_fp16 = mul(x = var_2643_cast_fp16, y = _inversed_2645_y_0_to_fp16)[name = string("_inversed_2645_cast_fp16")];
             bool attention_scores_67_transpose_x_0 = const()[name = string("attention_scores_67_transpose_x_0"), val = bool(false)];
             bool attention_scores_67_transpose_y_0 = const()[name = string("attention_scores_67_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_67_cast_fp16 = matmul(transpose_x = attention_scores_67_transpose_x_0, transpose_y = attention_scores_67_transpose_y_0, x = query_layer_69_cast_fp16, y = _inversed_2645_cast_fp16)[name = string("attention_scores_67_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_67_cast_fp16 = matmul(transpose_x = attention_scores_67_transpose_x_0, transpose_y = attention_scores_67_transpose_y_0, x = query_layer_69_cast_fp16, y = _inversed_2645_cast_fp16)[name = string("attention_scores_67_cast_fp16")];
             bool c2p_att_45_transpose_x_0 = const()[name = string("c2p_att_45_transpose_x_0"), val = bool(false)];
             bool c2p_att_45_transpose_y_0 = const()[name = string("c2p_att_45_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2701_to_fp16 = const()[name = string("op_2701_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569716416)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_45_cast_fp16 = matmul(transpose_x = c2p_att_45_transpose_x_0, transpose_y = c2p_att_45_transpose_y_0, x = query_layer_69_cast_fp16, y = var_2701_to_fp16)[name = string("c2p_att_45_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_45_cast_fp16 = matmul(transpose_x = c2p_att_45_transpose_x_0, transpose_y = c2p_att_45_transpose_y_0, x = query_layer_69_cast_fp16, y = var_2701_to_fp16)[name = string("c2p_att_45_cast_fp16")];
             bool c2p_att_47_validate_indices_0 = const()[name = string("c2p_att_47_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_47_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_47_validate_indices_0, x = c2p_att_45_cast_fp16)[name = string("c2p_att_47_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_47_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_47_validate_indices_0, x = c2p_att_45_cast_fp16)[name = string("c2p_att_47_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2714_y_0_to_fp16 = const()[name = string("_inversed_2714_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_2714_cast_fp16 = mul(x = c2p_att_47_cast_fp16_cast_int16, y = _inversed_2714_y_0_to_fp16)[name = string("_inversed_2714_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_2714_cast_fp16 = mul(x = c2p_att_47_cast_fp16_cast_int16, y = _inversed_2714_y_0_to_fp16)[name = string("_inversed_2714_cast_fp16")];
             bool p2c_att_45_transpose_x_0 = const()[name = string("p2c_att_45_transpose_x_0"), val = bool(false)];
             bool p2c_att_45_transpose_y_0 = const()[name = string("p2c_att_45_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2729_to_fp16 = const()[name = string("op_2729_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570765056)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_45_cast_fp16 = matmul(transpose_x = p2c_att_45_transpose_x_0, transpose_y = p2c_att_45_transpose_y_0, x = key_layer_23_cast_fp16, y = var_2729_to_fp16)[name = string("p2c_att_45_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_45_cast_fp16 = matmul(transpose_x = p2c_att_45_transpose_x_0, transpose_y = p2c_att_45_transpose_y_0, x = key_layer_23_cast_fp16, y = var_2729_to_fp16)[name = string("p2c_att_45_cast_fp16")];
             bool var_2737_validate_indices_0 = const()[name = string("op_2737_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_2737_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2737_validate_indices_0, x = p2c_att_45_cast_fp16)[name = string("op_2737_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_2737_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2737_validate_indices_0, x = p2c_att_45_cast_fp16)[name = string("op_2737_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_47_perm_0 = const()[name = string("p2c_att_47_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2740_y_0_to_fp16 = const()[name = string("_inversed_2740_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_47_cast_fp16 = transpose(perm = p2c_att_47_perm_0, x = var_2737_cast_fp16_cast_int16)[name = string("transpose_73")];
-            tensor<fp16, [16, 512, 512]> _inversed_2740_cast_fp16 = mul(x = p2c_att_47_cast_fp16, y = _inversed_2740_y_0_to_fp16)[name = string("_inversed_2740_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_23_cast_fp16 = add(x = _inversed_2714_cast_fp16, y = _inversed_2740_cast_fp16)[name = string("rel_att_23_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_69_cast_fp16 = add(x = attention_scores_67_cast_fp16, y = rel_att_23_cast_fp16)[name = string("attention_scores_69_cast_fp16")];
-            tensor<int32, [4]> var_2745 = const()[name = string("op_2745"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_71_cast_fp16 = reshape(shape = var_2745, x = attention_scores_69_cast_fp16)[name = string("attention_scores_71_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_273_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_71_cast_fp16, cond = var_339)[name = string("input_273_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_275_cast_fp16 = softmax(axis = var_56, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")];
-            tensor<int32, [3]> var_2754 = const()[name = string("op_2754"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_2755_cast_fp16 = reshape(shape = var_2754, x = input_275_cast_fp16)[name = string("op_2755_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_47_cast_fp16 = transpose(perm = p2c_att_47_perm_0, x = var_2737_cast_fp16_cast_int16)[name = string("transpose_73")];
+            tensor<fp16, [16, 224, 224]> _inversed_2740_cast_fp16 = mul(x = p2c_att_47_cast_fp16, y = _inversed_2740_y_0_to_fp16)[name = string("_inversed_2740_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_23_cast_fp16 = add(x = _inversed_2714_cast_fp16, y = _inversed_2740_cast_fp16)[name = string("rel_att_23_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_69_cast_fp16 = add(x = attention_scores_67_cast_fp16, y = rel_att_23_cast_fp16)[name = string("attention_scores_69_cast_fp16")];
+            tensor<int32, [4]> var_2745 = const()[name = string("op_2745"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_71_cast_fp16 = reshape(shape = var_2745, x = attention_scores_69_cast_fp16)[name = string("attention_scores_71_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_273_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_71_cast_fp16, cond = var_339)[name = string("input_273_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_275_cast_fp16 = softmax(axis = var_56, x = input_273_cast_fp16)[name = string("input_275_cast_fp16")];
+            tensor<int32, [3]> var_2754 = const()[name = string("op_2754"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_2755_cast_fp16 = reshape(shape = var_2754, x = input_275_cast_fp16)[name = string("op_2755_cast_fp16")];
             bool context_layer_45_transpose_x_0 = const()[name = string("context_layer_45_transpose_x_0"), val = bool(false)];
             bool context_layer_45_transpose_y_0 = const()[name = string("context_layer_45_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = var_2755_cast_fp16, y = value_layer_23_cast_fp16)[name = string("context_layer_45_cast_fp16")];
-            tensor<int32, [4]> var_2759 = const()[name = string("op_2759"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2760_cast_fp16 = reshape(shape = var_2759, x = context_layer_45_cast_fp16)[name = string("op_2760_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_45_cast_fp16 = matmul(transpose_x = context_layer_45_transpose_x_0, transpose_y = context_layer_45_transpose_y_0, x = var_2755_cast_fp16, y = value_layer_23_cast_fp16)[name = string("context_layer_45_cast_fp16")];
+            tensor<int32, [4]> var_2759 = const()[name = string("op_2759"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2760_cast_fp16 = reshape(shape = var_2759, x = context_layer_45_cast_fp16)[name = string("op_2760_cast_fp16")];
             tensor<int32, [4]> var_2761 = const()[name = string("op_2761"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2766 = const()[name = string("op_2766"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_2762_cast_fp16 = transpose(perm = var_2761, x = var_2760_cast_fp16)[name = string("transpose_72")];
-            tensor<fp16, [1, 512, 1024]> input_277_cast_fp16 = reshape(shape = var_2766, x = var_2762_cast_fp16)[name = string("input_277_cast_fp16")];
+            tensor<int32, [3]> var_2766 = const()[name = string("op_2766"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_2762_cast_fp16 = transpose(perm = var_2761, x = var_2760_cast_fp16)[name = string("transpose_72")];
+            tensor<fp16, [1, 224, 1024]> input_277_cast_fp16 = reshape(shape = var_2766, x = var_2762_cast_fp16)[name = string("input_277_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_11_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(571813696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573910912)))];
-            tensor<fp16, [1, 512, 1024]> linear_93_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_277_cast_fp16)[name = string("linear_93_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_281_cast_fp16 = add(x = linear_93_cast_fp16, y = input_269_cast_fp16)[name = string("input_281_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_93_cast_fp16 = linear(bias = encoder_encoder_layer_11_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_attention_output_dense_weight_to_fp16, x = input_277_cast_fp16)[name = string("linear_93_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_281_cast_fp16 = add(x = linear_93_cast_fp16, y = input_269_cast_fp16)[name = string("input_281_cast_fp16")];
             tensor<int32, [1]> input_283_axes_0 = const()[name = string("input_283_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573913024)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573915136)))];
-            tensor<fp16, [1, 512, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_283_cast_fp16 = layer_norm(axes = input_283_axes_0, beta = encoder_encoder_layer_11_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_attention_output_LayerNorm_weight_to_fp16, x = input_281_cast_fp16)[name = string("input_283_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_11_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(573917248)))];
             tensor<fp16, [4096]> encoder_encoder_layer_11_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582305920)))];
-            tensor<fp16, [1, 512, 4096]> linear_94_cast_fp16 = linear(bias = encoder_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_94_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_94_cast_fp16 = linear(bias = encoder_encoder_layer_11_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_11_intermediate_dense_weight_to_fp16, x = input_283_cast_fp16)[name = string("linear_94_cast_fp16")];
             string input_287_mode_0 = const()[name = string("input_287_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_287_cast_fp16 = gelu(mode = input_287_mode_0, x = linear_94_cast_fp16)[name = string("input_287_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_287_cast_fp16 = gelu(mode = input_287_mode_0, x = linear_94_cast_fp16)[name = string("input_287_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_11_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582314176)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590702848)))];
-            tensor<fp16, [1, 512, 1024]> linear_95_cast_fp16 = linear(bias = encoder_encoder_layer_11_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_output_dense_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_95_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_291_cast_fp16 = add(x = linear_95_cast_fp16, y = input_283_cast_fp16)[name = string("input_291_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_95_cast_fp16 = linear(bias = encoder_encoder_layer_11_output_dense_bias_to_fp16, weight = encoder_encoder_layer_11_output_dense_weight_to_fp16, x = input_287_cast_fp16)[name = string("linear_95_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_291_cast_fp16 = add(x = linear_95_cast_fp16, y = input_283_cast_fp16)[name = string("input_291_cast_fp16")];
             tensor<int32, [1]> input_293_axes_0 = const()[name = string("input_293_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590704960)))];
             tensor<fp16, [1024]> encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590707072)))];
-            tensor<fp16, [1, 512, 1024]> input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_291_cast_fp16)[name = string("input_293_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_293_cast_fp16 = layer_norm(axes = input_293_axes_0, beta = encoder_encoder_layer_11_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_11_output_LayerNorm_weight_to_fp16, x = input_291_cast_fp16)[name = string("input_293_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(590709184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592806400)))];
-            tensor<fp16, [1, 512, 1024]> linear_96_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_96_cast_fp16")];
-            tensor<int32, [4]> var_2817 = const()[name = string("op_2817"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_243_cast_fp16 = reshape(shape = var_2817, x = linear_96_cast_fp16)[name = string("x_243_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_96_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_query_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_96_cast_fp16")];
+            tensor<int32, [4]> var_2817 = const()[name = string("op_2817"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_243_cast_fp16 = reshape(shape = var_2817, x = linear_96_cast_fp16)[name = string("x_243_cast_fp16")];
             tensor<int32, [4]> var_2819 = const()[name = string("op_2819"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2824 = const()[name = string("op_2824"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2820_cast_fp16 = transpose(perm = var_2819, x = x_243_cast_fp16)[name = string("transpose_71")];
-            tensor<fp16, [16, 512, 64]> query_layer_75_cast_fp16 = reshape(shape = var_2824, x = var_2820_cast_fp16)[name = string("query_layer_75_cast_fp16")];
+            tensor<int32, [3]> var_2824 = const()[name = string("op_2824"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2820_cast_fp16 = transpose(perm = var_2819, x = x_243_cast_fp16)[name = string("transpose_71")];
+            tensor<fp16, [16, 224, 64]> query_layer_75_cast_fp16 = reshape(shape = var_2824, x = var_2820_cast_fp16)[name = string("query_layer_75_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(592808512)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594905728)))];
-            tensor<fp16, [1, 512, 1024]> linear_97_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_97_cast_fp16")];
-            tensor<int32, [4]> var_2831 = const()[name = string("op_2831"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_247_cast_fp16 = reshape(shape = var_2831, x = linear_97_cast_fp16)[name = string("x_247_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_97_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_key_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_97_cast_fp16")];
+            tensor<int32, [4]> var_2831 = const()[name = string("op_2831"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_247_cast_fp16 = reshape(shape = var_2831, x = linear_97_cast_fp16)[name = string("x_247_cast_fp16")];
             tensor<int32, [4]> var_2833 = const()[name = string("op_2833"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2834_cast_fp16 = transpose(perm = var_2833, x = x_247_cast_fp16)[name = string("transpose_70")];
-            tensor<fp16, [16, 512, 64]> key_layer_25_cast_fp16 = reshape(shape = var_2838, x = var_2834_cast_fp16)[name = string("key_layer_25_cast_fp16")];
+            tensor<int32, [3]> var_2838 = const()[name = string("op_2838"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2834_cast_fp16 = transpose(perm = var_2833, x = x_247_cast_fp16)[name = string("transpose_70")];
+            tensor<fp16, [16, 224, 64]> key_layer_25_cast_fp16 = reshape(shape = var_2838, x = var_2834_cast_fp16)[name = string("key_layer_25_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594907840)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597005056)))];
-            tensor<fp16, [1, 512, 1024]> linear_98_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_98_cast_fp16")];
-            tensor<int32, [4]> var_2845 = const()[name = string("op_2845"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_251_cast_fp16 = reshape(shape = var_2845, x = linear_98_cast_fp16)[name = string("x_251_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_98_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_12_attention_self_value_proj_weight_to_fp16, x = input_293_cast_fp16)[name = string("linear_98_cast_fp16")];
+            tensor<int32, [4]> var_2845 = const()[name = string("op_2845"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_251_cast_fp16 = reshape(shape = var_2845, x = linear_98_cast_fp16)[name = string("x_251_cast_fp16")];
             tensor<int32, [4]> var_2847 = const()[name = string("op_2847"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2852 = const()[name = string("op_2852"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2848_cast_fp16 = transpose(perm = var_2847, x = x_251_cast_fp16)[name = string("transpose_69")];
-            tensor<fp16, [16, 512, 64]> value_layer_25_cast_fp16 = reshape(shape = var_2852, x = var_2848_cast_fp16)[name = string("value_layer_25_cast_fp16")];
+            tensor<int32, [3]> var_2852 = const()[name = string("op_2852"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2848_cast_fp16 = transpose(perm = var_2847, x = x_251_cast_fp16)[name = string("transpose_69")];
+            tensor<fp16, [16, 224, 64]> value_layer_25_cast_fp16 = reshape(shape = var_2852, x = var_2848_cast_fp16)[name = string("value_layer_25_cast_fp16")];
             tensor<int32, [3]> var_2862_perm_0 = const()[name = string("op_2862_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2864_y_0_to_fp16 = const()[name = string("_inversed_2864_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_2862_cast_fp16 = transpose(perm = var_2862_perm_0, x = key_layer_25_cast_fp16)[name = string("transpose_68")];
-            tensor<fp16, [16, 64, 512]> _inversed_2864_cast_fp16 = mul(x = var_2862_cast_fp16, y = _inversed_2864_y_0_to_fp16)[name = string("_inversed_2864_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_2862_cast_fp16 = transpose(perm = var_2862_perm_0, x = key_layer_25_cast_fp16)[name = string("transpose_68")];
+            tensor<fp16, [16, 64, 224]> _inversed_2864_cast_fp16 = mul(x = var_2862_cast_fp16, y = _inversed_2864_y_0_to_fp16)[name = string("_inversed_2864_cast_fp16")];
             bool attention_scores_73_transpose_x_0 = const()[name = string("attention_scores_73_transpose_x_0"), val = bool(false)];
             bool attention_scores_73_transpose_y_0 = const()[name = string("attention_scores_73_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_73_cast_fp16 = matmul(transpose_x = attention_scores_73_transpose_x_0, transpose_y = attention_scores_73_transpose_y_0, x = query_layer_75_cast_fp16, y = _inversed_2864_cast_fp16)[name = string("attention_scores_73_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_73_cast_fp16 = matmul(transpose_x = attention_scores_73_transpose_x_0, transpose_y = attention_scores_73_transpose_y_0, x = query_layer_75_cast_fp16, y = _inversed_2864_cast_fp16)[name = string("attention_scores_73_cast_fp16")];
             bool c2p_att_49_transpose_x_0 = const()[name = string("c2p_att_49_transpose_x_0"), val = bool(false)];
             bool c2p_att_49_transpose_y_0 = const()[name = string("c2p_att_49_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2920_to_fp16 = const()[name = string("op_2920_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597007168)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_49_cast_fp16 = matmul(transpose_x = c2p_att_49_transpose_x_0, transpose_y = c2p_att_49_transpose_y_0, x = query_layer_75_cast_fp16, y = var_2920_to_fp16)[name = string("c2p_att_49_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_49_cast_fp16 = matmul(transpose_x = c2p_att_49_transpose_x_0, transpose_y = c2p_att_49_transpose_y_0, x = query_layer_75_cast_fp16, y = var_2920_to_fp16)[name = string("c2p_att_49_cast_fp16")];
             bool c2p_att_51_validate_indices_0 = const()[name = string("c2p_att_51_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_51_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_51_validate_indices_0, x = c2p_att_49_cast_fp16)[name = string("c2p_att_51_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_51_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_51_validate_indices_0, x = c2p_att_49_cast_fp16)[name = string("c2p_att_51_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_2933_y_0_to_fp16 = const()[name = string("_inversed_2933_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_2933_cast_fp16 = mul(x = c2p_att_51_cast_fp16_cast_int16, y = _inversed_2933_y_0_to_fp16)[name = string("_inversed_2933_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_2933_cast_fp16 = mul(x = c2p_att_51_cast_fp16_cast_int16, y = _inversed_2933_y_0_to_fp16)[name = string("_inversed_2933_cast_fp16")];
             bool p2c_att_49_transpose_x_0 = const()[name = string("p2c_att_49_transpose_x_0"), val = bool(false)];
             bool p2c_att_49_transpose_y_0 = const()[name = string("p2c_att_49_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_2948_to_fp16 = const()[name = string("op_2948_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598055808)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_49_cast_fp16 = matmul(transpose_x = p2c_att_49_transpose_x_0, transpose_y = p2c_att_49_transpose_y_0, x = key_layer_25_cast_fp16, y = var_2948_to_fp16)[name = string("p2c_att_49_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_49_cast_fp16 = matmul(transpose_x = p2c_att_49_transpose_x_0, transpose_y = p2c_att_49_transpose_y_0, x = key_layer_25_cast_fp16, y = var_2948_to_fp16)[name = string("p2c_att_49_cast_fp16")];
             bool var_2956_validate_indices_0 = const()[name = string("op_2956_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_2956_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2956_validate_indices_0, x = p2c_att_49_cast_fp16)[name = string("op_2956_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_2956_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_2956_validate_indices_0, x = p2c_att_49_cast_fp16)[name = string("op_2956_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_51_perm_0 = const()[name = string("p2c_att_51_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_2959_y_0_to_fp16 = const()[name = string("_inversed_2959_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_51_cast_fp16 = transpose(perm = p2c_att_51_perm_0, x = var_2956_cast_fp16_cast_int16)[name = string("transpose_67")];
-            tensor<fp16, [16, 512, 512]> _inversed_2959_cast_fp16 = mul(x = p2c_att_51_cast_fp16, y = _inversed_2959_y_0_to_fp16)[name = string("_inversed_2959_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_25_cast_fp16 = add(x = _inversed_2933_cast_fp16, y = _inversed_2959_cast_fp16)[name = string("rel_att_25_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_75_cast_fp16 = add(x = attention_scores_73_cast_fp16, y = rel_att_25_cast_fp16)[name = string("attention_scores_75_cast_fp16")];
-            tensor<int32, [4]> var_2964 = const()[name = string("op_2964"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_77_cast_fp16 = reshape(shape = var_2964, x = attention_scores_75_cast_fp16)[name = string("attention_scores_77_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_297_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_77_cast_fp16, cond = var_339)[name = string("input_297_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_299_cast_fp16 = softmax(axis = var_56, x = input_297_cast_fp16)[name = string("input_299_cast_fp16")];
-            tensor<int32, [3]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_2974_cast_fp16 = reshape(shape = var_2973, x = input_299_cast_fp16)[name = string("op_2974_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_51_cast_fp16 = transpose(perm = p2c_att_51_perm_0, x = var_2956_cast_fp16_cast_int16)[name = string("transpose_67")];
+            tensor<fp16, [16, 224, 224]> _inversed_2959_cast_fp16 = mul(x = p2c_att_51_cast_fp16, y = _inversed_2959_y_0_to_fp16)[name = string("_inversed_2959_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_25_cast_fp16 = add(x = _inversed_2933_cast_fp16, y = _inversed_2959_cast_fp16)[name = string("rel_att_25_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_75_cast_fp16 = add(x = attention_scores_73_cast_fp16, y = rel_att_25_cast_fp16)[name = string("attention_scores_75_cast_fp16")];
+            tensor<int32, [4]> var_2964 = const()[name = string("op_2964"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_77_cast_fp16 = reshape(shape = var_2964, x = attention_scores_75_cast_fp16)[name = string("attention_scores_77_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_297_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_77_cast_fp16, cond = var_339)[name = string("input_297_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_299_cast_fp16 = softmax(axis = var_56, x = input_297_cast_fp16)[name = string("input_299_cast_fp16")];
+            tensor<int32, [3]> var_2973 = const()[name = string("op_2973"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_2974_cast_fp16 = reshape(shape = var_2973, x = input_299_cast_fp16)[name = string("op_2974_cast_fp16")];
             bool context_layer_49_transpose_x_0 = const()[name = string("context_layer_49_transpose_x_0"), val = bool(false)];
             bool context_layer_49_transpose_y_0 = const()[name = string("context_layer_49_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_49_cast_fp16 = matmul(transpose_x = context_layer_49_transpose_x_0, transpose_y = context_layer_49_transpose_y_0, x = var_2974_cast_fp16, y = value_layer_25_cast_fp16)[name = string("context_layer_49_cast_fp16")];
-            tensor<int32, [4]> var_2978 = const()[name = string("op_2978"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_2979_cast_fp16 = reshape(shape = var_2978, x = context_layer_49_cast_fp16)[name = string("op_2979_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_49_cast_fp16 = matmul(transpose_x = context_layer_49_transpose_x_0, transpose_y = context_layer_49_transpose_y_0, x = var_2974_cast_fp16, y = value_layer_25_cast_fp16)[name = string("context_layer_49_cast_fp16")];
+            tensor<int32, [4]> var_2978 = const()[name = string("op_2978"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_2979_cast_fp16 = reshape(shape = var_2978, x = context_layer_49_cast_fp16)[name = string("op_2979_cast_fp16")];
             tensor<int32, [4]> var_2980 = const()[name = string("op_2980"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_2985 = const()[name = string("op_2985"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_2981_cast_fp16 = transpose(perm = var_2980, x = var_2979_cast_fp16)[name = string("transpose_66")];
-            tensor<fp16, [1, 512, 1024]> input_301_cast_fp16 = reshape(shape = var_2985, x = var_2981_cast_fp16)[name = string("input_301_cast_fp16")];
+            tensor<int32, [3]> var_2985 = const()[name = string("op_2985"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_2981_cast_fp16 = transpose(perm = var_2980, x = var_2979_cast_fp16)[name = string("transpose_66")];
+            tensor<fp16, [1, 224, 1024]> input_301_cast_fp16 = reshape(shape = var_2985, x = var_2981_cast_fp16)[name = string("input_301_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_12_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599104448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601201664)))];
-            tensor<fp16, [1, 512, 1024]> linear_101_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_attention_output_dense_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_101_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_305_cast_fp16 = add(x = linear_101_cast_fp16, y = input_293_cast_fp16)[name = string("input_305_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_101_cast_fp16 = linear(bias = encoder_encoder_layer_12_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_attention_output_dense_weight_to_fp16, x = input_301_cast_fp16)[name = string("linear_101_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_305_cast_fp16 = add(x = linear_101_cast_fp16, y = input_293_cast_fp16)[name = string("input_305_cast_fp16")];
             tensor<int32, [1]> input_307_axes_0 = const()[name = string("input_307_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601203776)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601205888)))];
-            tensor<fp16, [1, 512, 1024]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_307_cast_fp16 = layer_norm(axes = input_307_axes_0, beta = encoder_encoder_layer_12_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_attention_output_LayerNorm_weight_to_fp16, x = input_305_cast_fp16)[name = string("input_307_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_12_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601208000)))];
             tensor<fp16, [4096]> encoder_encoder_layer_12_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609596672)))];
-            tensor<fp16, [1, 512, 4096]> linear_102_cast_fp16 = linear(bias = encoder_encoder_layer_12_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_12_intermediate_dense_weight_to_fp16, x = input_307_cast_fp16)[name = string("linear_102_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_102_cast_fp16 = linear(bias = encoder_encoder_layer_12_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_12_intermediate_dense_weight_to_fp16, x = input_307_cast_fp16)[name = string("linear_102_cast_fp16")];
             string input_311_mode_0 = const()[name = string("input_311_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = linear_102_cast_fp16)[name = string("input_311_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_311_cast_fp16 = gelu(mode = input_311_mode_0, x = linear_102_cast_fp16)[name = string("input_311_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_12_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(609604928)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617993600)))];
-            tensor<fp16, [1, 512, 1024]> linear_103_cast_fp16 = linear(bias = encoder_encoder_layer_12_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_output_dense_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_103_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_315_cast_fp16 = add(x = linear_103_cast_fp16, y = input_307_cast_fp16)[name = string("input_315_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_103_cast_fp16 = linear(bias = encoder_encoder_layer_12_output_dense_bias_to_fp16, weight = encoder_encoder_layer_12_output_dense_weight_to_fp16, x = input_311_cast_fp16)[name = string("linear_103_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_315_cast_fp16 = add(x = linear_103_cast_fp16, y = input_307_cast_fp16)[name = string("input_315_cast_fp16")];
             tensor<int32, [1]> input_317_axes_0 = const()[name = string("input_317_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617995712)))];
             tensor<fp16, [1024]> encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617997824)))];
-            tensor<fp16, [1, 512, 1024]> input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16, x = input_315_cast_fp16)[name = string("input_317_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_317_cast_fp16 = layer_norm(axes = input_317_axes_0, beta = encoder_encoder_layer_12_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_12_output_LayerNorm_weight_to_fp16, x = input_315_cast_fp16)[name = string("input_317_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(617999936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620097152)))];
-            tensor<fp16, [1, 512, 1024]> linear_104_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_104_cast_fp16")];
-            tensor<int32, [4]> var_3036 = const()[name = string("op_3036"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_263_cast_fp16 = reshape(shape = var_3036, x = linear_104_cast_fp16)[name = string("x_263_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_104_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_query_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_104_cast_fp16")];
+            tensor<int32, [4]> var_3036 = const()[name = string("op_3036"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_263_cast_fp16 = reshape(shape = var_3036, x = linear_104_cast_fp16)[name = string("x_263_cast_fp16")];
             tensor<int32, [4]> var_3038 = const()[name = string("op_3038"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3043 = const()[name = string("op_3043"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3039_cast_fp16 = transpose(perm = var_3038, x = x_263_cast_fp16)[name = string("transpose_65")];
-            tensor<fp16, [16, 512, 64]> query_layer_81_cast_fp16 = reshape(shape = var_3043, x = var_3039_cast_fp16)[name = string("query_layer_81_cast_fp16")];
+            tensor<int32, [3]> var_3043 = const()[name = string("op_3043"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3039_cast_fp16 = transpose(perm = var_3038, x = x_263_cast_fp16)[name = string("transpose_65")];
+            tensor<fp16, [16, 224, 64]> query_layer_81_cast_fp16 = reshape(shape = var_3043, x = var_3039_cast_fp16)[name = string("query_layer_81_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620099264)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622196480)))];
-            tensor<fp16, [1, 512, 1024]> linear_105_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_105_cast_fp16")];
-            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_267_cast_fp16 = reshape(shape = var_3050, x = linear_105_cast_fp16)[name = string("x_267_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_105_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_key_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_105_cast_fp16")];
+            tensor<int32, [4]> var_3050 = const()[name = string("op_3050"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_267_cast_fp16 = reshape(shape = var_3050, x = linear_105_cast_fp16)[name = string("x_267_cast_fp16")];
             tensor<int32, [4]> var_3052 = const()[name = string("op_3052"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3053_cast_fp16 = transpose(perm = var_3052, x = x_267_cast_fp16)[name = string("transpose_64")];
-            tensor<fp16, [16, 512, 64]> key_layer_27_cast_fp16 = reshape(shape = var_3057, x = var_3053_cast_fp16)[name = string("key_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3053_cast_fp16 = transpose(perm = var_3052, x = x_267_cast_fp16)[name = string("transpose_64")];
+            tensor<fp16, [16, 224, 64]> key_layer_27_cast_fp16 = reshape(shape = var_3057, x = var_3053_cast_fp16)[name = string("key_layer_27_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(622198592)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624295808)))];
-            tensor<fp16, [1, 512, 1024]> linear_106_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_106_cast_fp16")];
-            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_271_cast_fp16 = reshape(shape = var_3064, x = linear_106_cast_fp16)[name = string("x_271_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_106_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_13_attention_self_value_proj_weight_to_fp16, x = input_317_cast_fp16)[name = string("linear_106_cast_fp16")];
+            tensor<int32, [4]> var_3064 = const()[name = string("op_3064"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_271_cast_fp16 = reshape(shape = var_3064, x = linear_106_cast_fp16)[name = string("x_271_cast_fp16")];
             tensor<int32, [4]> var_3066 = const()[name = string("op_3066"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3071 = const()[name = string("op_3071"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3067_cast_fp16 = transpose(perm = var_3066, x = x_271_cast_fp16)[name = string("transpose_63")];
-            tensor<fp16, [16, 512, 64]> value_layer_27_cast_fp16 = reshape(shape = var_3071, x = var_3067_cast_fp16)[name = string("value_layer_27_cast_fp16")];
+            tensor<int32, [3]> var_3071 = const()[name = string("op_3071"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3067_cast_fp16 = transpose(perm = var_3066, x = x_271_cast_fp16)[name = string("transpose_63")];
+            tensor<fp16, [16, 224, 64]> value_layer_27_cast_fp16 = reshape(shape = var_3071, x = var_3067_cast_fp16)[name = string("value_layer_27_cast_fp16")];
             tensor<int32, [3]> var_3081_perm_0 = const()[name = string("op_3081_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3083_y_0_to_fp16 = const()[name = string("_inversed_3083_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_3081_cast_fp16 = transpose(perm = var_3081_perm_0, x = key_layer_27_cast_fp16)[name = string("transpose_62")];
-            tensor<fp16, [16, 64, 512]> _inversed_3083_cast_fp16 = mul(x = var_3081_cast_fp16, y = _inversed_3083_y_0_to_fp16)[name = string("_inversed_3083_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_3081_cast_fp16 = transpose(perm = var_3081_perm_0, x = key_layer_27_cast_fp16)[name = string("transpose_62")];
+            tensor<fp16, [16, 64, 224]> _inversed_3083_cast_fp16 = mul(x = var_3081_cast_fp16, y = _inversed_3083_y_0_to_fp16)[name = string("_inversed_3083_cast_fp16")];
             bool attention_scores_79_transpose_x_0 = const()[name = string("attention_scores_79_transpose_x_0"), val = bool(false)];
             bool attention_scores_79_transpose_y_0 = const()[name = string("attention_scores_79_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_79_cast_fp16 = matmul(transpose_x = attention_scores_79_transpose_x_0, transpose_y = attention_scores_79_transpose_y_0, x = query_layer_81_cast_fp16, y = _inversed_3083_cast_fp16)[name = string("attention_scores_79_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_79_cast_fp16 = matmul(transpose_x = attention_scores_79_transpose_x_0, transpose_y = attention_scores_79_transpose_y_0, x = query_layer_81_cast_fp16, y = _inversed_3083_cast_fp16)[name = string("attention_scores_79_cast_fp16")];
             bool c2p_att_53_transpose_x_0 = const()[name = string("c2p_att_53_transpose_x_0"), val = bool(false)];
             bool c2p_att_53_transpose_y_0 = const()[name = string("c2p_att_53_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3139_to_fp16 = const()[name = string("op_3139_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(624297920)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_53_cast_fp16 = matmul(transpose_x = c2p_att_53_transpose_x_0, transpose_y = c2p_att_53_transpose_y_0, x = query_layer_81_cast_fp16, y = var_3139_to_fp16)[name = string("c2p_att_53_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_53_cast_fp16 = matmul(transpose_x = c2p_att_53_transpose_x_0, transpose_y = c2p_att_53_transpose_y_0, x = query_layer_81_cast_fp16, y = var_3139_to_fp16)[name = string("c2p_att_53_cast_fp16")];
             bool c2p_att_55_validate_indices_0 = const()[name = string("c2p_att_55_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_55_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_55_validate_indices_0, x = c2p_att_53_cast_fp16)[name = string("c2p_att_55_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_55_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_55_validate_indices_0, x = c2p_att_53_cast_fp16)[name = string("c2p_att_55_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3152_y_0_to_fp16 = const()[name = string("_inversed_3152_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_3152_cast_fp16 = mul(x = c2p_att_55_cast_fp16_cast_int16, y = _inversed_3152_y_0_to_fp16)[name = string("_inversed_3152_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_3152_cast_fp16 = mul(x = c2p_att_55_cast_fp16_cast_int16, y = _inversed_3152_y_0_to_fp16)[name = string("_inversed_3152_cast_fp16")];
             bool p2c_att_53_transpose_x_0 = const()[name = string("p2c_att_53_transpose_x_0"), val = bool(false)];
             bool p2c_att_53_transpose_y_0 = const()[name = string("p2c_att_53_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3167_to_fp16 = const()[name = string("op_3167_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(625346560)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_53_cast_fp16 = matmul(transpose_x = p2c_att_53_transpose_x_0, transpose_y = p2c_att_53_transpose_y_0, x = key_layer_27_cast_fp16, y = var_3167_to_fp16)[name = string("p2c_att_53_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_53_cast_fp16 = matmul(transpose_x = p2c_att_53_transpose_x_0, transpose_y = p2c_att_53_transpose_y_0, x = key_layer_27_cast_fp16, y = var_3167_to_fp16)[name = string("p2c_att_53_cast_fp16")];
             bool var_3175_validate_indices_0 = const()[name = string("op_3175_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_3175_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3175_validate_indices_0, x = p2c_att_53_cast_fp16)[name = string("op_3175_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_3175_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3175_validate_indices_0, x = p2c_att_53_cast_fp16)[name = string("op_3175_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_55_perm_0 = const()[name = string("p2c_att_55_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3178_y_0_to_fp16 = const()[name = string("_inversed_3178_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_55_cast_fp16 = transpose(perm = p2c_att_55_perm_0, x = var_3175_cast_fp16_cast_int16)[name = string("transpose_61")];
-            tensor<fp16, [16, 512, 512]> _inversed_3178_cast_fp16 = mul(x = p2c_att_55_cast_fp16, y = _inversed_3178_y_0_to_fp16)[name = string("_inversed_3178_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_27_cast_fp16 = add(x = _inversed_3152_cast_fp16, y = _inversed_3178_cast_fp16)[name = string("rel_att_27_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_81_cast_fp16 = add(x = attention_scores_79_cast_fp16, y = rel_att_27_cast_fp16)[name = string("attention_scores_81_cast_fp16")];
-            tensor<int32, [4]> var_3183 = const()[name = string("op_3183"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_83_cast_fp16 = reshape(shape = var_3183, x = attention_scores_81_cast_fp16)[name = string("attention_scores_83_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_321_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_83_cast_fp16, cond = var_339)[name = string("input_321_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_323_cast_fp16 = softmax(axis = var_56, x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
-            tensor<int32, [3]> var_3192 = const()[name = string("op_3192"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_3193_cast_fp16 = reshape(shape = var_3192, x = input_323_cast_fp16)[name = string("op_3193_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_55_cast_fp16 = transpose(perm = p2c_att_55_perm_0, x = var_3175_cast_fp16_cast_int16)[name = string("transpose_61")];
+            tensor<fp16, [16, 224, 224]> _inversed_3178_cast_fp16 = mul(x = p2c_att_55_cast_fp16, y = _inversed_3178_y_0_to_fp16)[name = string("_inversed_3178_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_27_cast_fp16 = add(x = _inversed_3152_cast_fp16, y = _inversed_3178_cast_fp16)[name = string("rel_att_27_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_81_cast_fp16 = add(x = attention_scores_79_cast_fp16, y = rel_att_27_cast_fp16)[name = string("attention_scores_81_cast_fp16")];
+            tensor<int32, [4]> var_3183 = const()[name = string("op_3183"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_83_cast_fp16 = reshape(shape = var_3183, x = attention_scores_81_cast_fp16)[name = string("attention_scores_83_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_321_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_83_cast_fp16, cond = var_339)[name = string("input_321_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_323_cast_fp16 = softmax(axis = var_56, x = input_321_cast_fp16)[name = string("input_323_cast_fp16")];
+            tensor<int32, [3]> var_3192 = const()[name = string("op_3192"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_3193_cast_fp16 = reshape(shape = var_3192, x = input_323_cast_fp16)[name = string("op_3193_cast_fp16")];
             bool context_layer_53_transpose_x_0 = const()[name = string("context_layer_53_transpose_x_0"), val = bool(false)];
             bool context_layer_53_transpose_y_0 = const()[name = string("context_layer_53_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_53_cast_fp16 = matmul(transpose_x = context_layer_53_transpose_x_0, transpose_y = context_layer_53_transpose_y_0, x = var_3193_cast_fp16, y = value_layer_27_cast_fp16)[name = string("context_layer_53_cast_fp16")];
-            tensor<int32, [4]> var_3197 = const()[name = string("op_3197"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3198_cast_fp16 = reshape(shape = var_3197, x = context_layer_53_cast_fp16)[name = string("op_3198_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_53_cast_fp16 = matmul(transpose_x = context_layer_53_transpose_x_0, transpose_y = context_layer_53_transpose_y_0, x = var_3193_cast_fp16, y = value_layer_27_cast_fp16)[name = string("context_layer_53_cast_fp16")];
+            tensor<int32, [4]> var_3197 = const()[name = string("op_3197"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3198_cast_fp16 = reshape(shape = var_3197, x = context_layer_53_cast_fp16)[name = string("op_3198_cast_fp16")];
             tensor<int32, [4]> var_3199 = const()[name = string("op_3199"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3204 = const()[name = string("op_3204"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_3200_cast_fp16 = transpose(perm = var_3199, x = var_3198_cast_fp16)[name = string("transpose_60")];
-            tensor<fp16, [1, 512, 1024]> input_325_cast_fp16 = reshape(shape = var_3204, x = var_3200_cast_fp16)[name = string("input_325_cast_fp16")];
+            tensor<int32, [3]> var_3204 = const()[name = string("op_3204"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_3200_cast_fp16 = transpose(perm = var_3199, x = var_3198_cast_fp16)[name = string("transpose_60")];
+            tensor<fp16, [1, 224, 1024]> input_325_cast_fp16 = reshape(shape = var_3204, x = var_3200_cast_fp16)[name = string("input_325_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_13_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626395200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628492416)))];
-            tensor<fp16, [1, 512, 1024]> linear_109_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_attention_output_dense_weight_to_fp16, x = input_325_cast_fp16)[name = string("linear_109_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_329_cast_fp16 = add(x = linear_109_cast_fp16, y = input_317_cast_fp16)[name = string("input_329_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_109_cast_fp16 = linear(bias = encoder_encoder_layer_13_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_attention_output_dense_weight_to_fp16, x = input_325_cast_fp16)[name = string("linear_109_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_329_cast_fp16 = add(x = linear_109_cast_fp16, y = input_317_cast_fp16)[name = string("input_329_cast_fp16")];
             tensor<int32, [1]> input_331_axes_0 = const()[name = string("input_331_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628494528)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628496640)))];
-            tensor<fp16, [1, 512, 1024]> input_331_cast_fp16 = layer_norm(axes = input_331_axes_0, beta = encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16, x = input_329_cast_fp16)[name = string("input_331_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_331_cast_fp16 = layer_norm(axes = input_331_axes_0, beta = encoder_encoder_layer_13_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_attention_output_LayerNorm_weight_to_fp16, x = input_329_cast_fp16)[name = string("input_331_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_13_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628498752)))];
             tensor<fp16, [4096]> encoder_encoder_layer_13_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636887424)))];
-            tensor<fp16, [1, 512, 4096]> linear_110_cast_fp16 = linear(bias = encoder_encoder_layer_13_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_13_intermediate_dense_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_110_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_110_cast_fp16 = linear(bias = encoder_encoder_layer_13_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_13_intermediate_dense_weight_to_fp16, x = input_331_cast_fp16)[name = string("linear_110_cast_fp16")];
             string input_335_mode_0 = const()[name = string("input_335_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_335_cast_fp16 = gelu(mode = input_335_mode_0, x = linear_110_cast_fp16)[name = string("input_335_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_335_cast_fp16 = gelu(mode = input_335_mode_0, x = linear_110_cast_fp16)[name = string("input_335_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_13_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(636895680)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645284352)))];
-            tensor<fp16, [1, 512, 1024]> linear_111_cast_fp16 = linear(bias = encoder_encoder_layer_13_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_output_dense_weight_to_fp16, x = input_335_cast_fp16)[name = string("linear_111_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_339_cast_fp16 = add(x = linear_111_cast_fp16, y = input_331_cast_fp16)[name = string("input_339_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_111_cast_fp16 = linear(bias = encoder_encoder_layer_13_output_dense_bias_to_fp16, weight = encoder_encoder_layer_13_output_dense_weight_to_fp16, x = input_335_cast_fp16)[name = string("linear_111_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_339_cast_fp16 = add(x = linear_111_cast_fp16, y = input_331_cast_fp16)[name = string("input_339_cast_fp16")];
             tensor<int32, [1]> input_341_axes_0 = const()[name = string("input_341_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645286464)))];
             tensor<fp16, [1024]> encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645288576)))];
-            tensor<fp16, [1, 512, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_0, beta = encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_341_cast_fp16 = layer_norm(axes = input_341_axes_0, beta = encoder_encoder_layer_13_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_13_output_LayerNorm_weight_to_fp16, x = input_339_cast_fp16)[name = string("input_341_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645290688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647387904)))];
-            tensor<fp16, [1, 512, 1024]> linear_112_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_112_cast_fp16")];
-            tensor<int32, [4]> var_3255 = const()[name = string("op_3255"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_283_cast_fp16 = reshape(shape = var_3255, x = linear_112_cast_fp16)[name = string("x_283_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_112_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_query_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_112_cast_fp16")];
+            tensor<int32, [4]> var_3255 = const()[name = string("op_3255"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_283_cast_fp16 = reshape(shape = var_3255, x = linear_112_cast_fp16)[name = string("x_283_cast_fp16")];
             tensor<int32, [4]> var_3257 = const()[name = string("op_3257"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3262 = const()[name = string("op_3262"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3258_cast_fp16 = transpose(perm = var_3257, x = x_283_cast_fp16)[name = string("transpose_59")];
-            tensor<fp16, [16, 512, 64]> query_layer_87_cast_fp16 = reshape(shape = var_3262, x = var_3258_cast_fp16)[name = string("query_layer_87_cast_fp16")];
+            tensor<int32, [3]> var_3262 = const()[name = string("op_3262"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3258_cast_fp16 = transpose(perm = var_3257, x = x_283_cast_fp16)[name = string("transpose_59")];
+            tensor<fp16, [16, 224, 64]> query_layer_87_cast_fp16 = reshape(shape = var_3262, x = var_3258_cast_fp16)[name = string("query_layer_87_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(647390016)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(649487232)))];
-            tensor<fp16, [1, 512, 1024]> linear_113_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_113_cast_fp16")];
-            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_287_cast_fp16 = reshape(shape = var_3269, x = linear_113_cast_fp16)[name = string("x_287_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_113_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_key_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_113_cast_fp16")];
+            tensor<int32, [4]> var_3269 = const()[name = string("op_3269"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_287_cast_fp16 = reshape(shape = var_3269, x = linear_113_cast_fp16)[name = string("x_287_cast_fp16")];
             tensor<int32, [4]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3276 = const()[name = string("op_3276"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3272_cast_fp16 = transpose(perm = var_3271, x = x_287_cast_fp16)[name = string("transpose_58")];
-            tensor<fp16, [16, 512, 64]> key_layer_29_cast_fp16 = reshape(shape = var_3276, x = var_3272_cast_fp16)[name = string("key_layer_29_cast_fp16")];
+            tensor<int32, [3]> var_3276 = const()[name = string("op_3276"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3272_cast_fp16 = transpose(perm = var_3271, x = x_287_cast_fp16)[name = string("transpose_58")];
+            tensor<fp16, [16, 224, 64]> key_layer_29_cast_fp16 = reshape(shape = var_3276, x = var_3272_cast_fp16)[name = string("key_layer_29_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(649489344)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651586560)))];
-            tensor<fp16, [1, 512, 1024]> linear_114_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_114_cast_fp16")];
-            tensor<int32, [4]> var_3283 = const()[name = string("op_3283"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_291_cast_fp16 = reshape(shape = var_3283, x = linear_114_cast_fp16)[name = string("x_291_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_114_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_14_attention_self_value_proj_weight_to_fp16, x = input_341_cast_fp16)[name = string("linear_114_cast_fp16")];
+            tensor<int32, [4]> var_3283 = const()[name = string("op_3283"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_291_cast_fp16 = reshape(shape = var_3283, x = linear_114_cast_fp16)[name = string("x_291_cast_fp16")];
             tensor<int32, [4]> var_3285 = const()[name = string("op_3285"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3286_cast_fp16 = transpose(perm = var_3285, x = x_291_cast_fp16)[name = string("transpose_57")];
-            tensor<fp16, [16, 512, 64]> value_layer_29_cast_fp16 = reshape(shape = var_3290, x = var_3286_cast_fp16)[name = string("value_layer_29_cast_fp16")];
+            tensor<int32, [3]> var_3290 = const()[name = string("op_3290"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3286_cast_fp16 = transpose(perm = var_3285, x = x_291_cast_fp16)[name = string("transpose_57")];
+            tensor<fp16, [16, 224, 64]> value_layer_29_cast_fp16 = reshape(shape = var_3290, x = var_3286_cast_fp16)[name = string("value_layer_29_cast_fp16")];
             tensor<int32, [3]> var_3300_perm_0 = const()[name = string("op_3300_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3302_y_0_to_fp16 = const()[name = string("_inversed_3302_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_3300_cast_fp16 = transpose(perm = var_3300_perm_0, x = key_layer_29_cast_fp16)[name = string("transpose_56")];
-            tensor<fp16, [16, 64, 512]> _inversed_3302_cast_fp16 = mul(x = var_3300_cast_fp16, y = _inversed_3302_y_0_to_fp16)[name = string("_inversed_3302_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_3300_cast_fp16 = transpose(perm = var_3300_perm_0, x = key_layer_29_cast_fp16)[name = string("transpose_56")];
+            tensor<fp16, [16, 64, 224]> _inversed_3302_cast_fp16 = mul(x = var_3300_cast_fp16, y = _inversed_3302_y_0_to_fp16)[name = string("_inversed_3302_cast_fp16")];
             bool attention_scores_85_transpose_x_0 = const()[name = string("attention_scores_85_transpose_x_0"), val = bool(false)];
             bool attention_scores_85_transpose_y_0 = const()[name = string("attention_scores_85_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_85_cast_fp16 = matmul(transpose_x = attention_scores_85_transpose_x_0, transpose_y = attention_scores_85_transpose_y_0, x = query_layer_87_cast_fp16, y = _inversed_3302_cast_fp16)[name = string("attention_scores_85_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_85_cast_fp16 = matmul(transpose_x = attention_scores_85_transpose_x_0, transpose_y = attention_scores_85_transpose_y_0, x = query_layer_87_cast_fp16, y = _inversed_3302_cast_fp16)[name = string("attention_scores_85_cast_fp16")];
             bool c2p_att_57_transpose_x_0 = const()[name = string("c2p_att_57_transpose_x_0"), val = bool(false)];
             bool c2p_att_57_transpose_y_0 = const()[name = string("c2p_att_57_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3358_to_fp16 = const()[name = string("op_3358_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651588672)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_57_cast_fp16 = matmul(transpose_x = c2p_att_57_transpose_x_0, transpose_y = c2p_att_57_transpose_y_0, x = query_layer_87_cast_fp16, y = var_3358_to_fp16)[name = string("c2p_att_57_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_57_cast_fp16 = matmul(transpose_x = c2p_att_57_transpose_x_0, transpose_y = c2p_att_57_transpose_y_0, x = query_layer_87_cast_fp16, y = var_3358_to_fp16)[name = string("c2p_att_57_cast_fp16")];
             bool c2p_att_59_validate_indices_0 = const()[name = string("c2p_att_59_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_59_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_59_validate_indices_0, x = c2p_att_57_cast_fp16)[name = string("c2p_att_59_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_59_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_59_validate_indices_0, x = c2p_att_57_cast_fp16)[name = string("c2p_att_59_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3371_y_0_to_fp16 = const()[name = string("_inversed_3371_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_3371_cast_fp16 = mul(x = c2p_att_59_cast_fp16_cast_int16, y = _inversed_3371_y_0_to_fp16)[name = string("_inversed_3371_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_3371_cast_fp16 = mul(x = c2p_att_59_cast_fp16_cast_int16, y = _inversed_3371_y_0_to_fp16)[name = string("_inversed_3371_cast_fp16")];
             bool p2c_att_57_transpose_x_0 = const()[name = string("p2c_att_57_transpose_x_0"), val = bool(false)];
             bool p2c_att_57_transpose_y_0 = const()[name = string("p2c_att_57_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3386_to_fp16 = const()[name = string("op_3386_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(652637312)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_57_cast_fp16 = matmul(transpose_x = p2c_att_57_transpose_x_0, transpose_y = p2c_att_57_transpose_y_0, x = key_layer_29_cast_fp16, y = var_3386_to_fp16)[name = string("p2c_att_57_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_57_cast_fp16 = matmul(transpose_x = p2c_att_57_transpose_x_0, transpose_y = p2c_att_57_transpose_y_0, x = key_layer_29_cast_fp16, y = var_3386_to_fp16)[name = string("p2c_att_57_cast_fp16")];
             bool var_3394_validate_indices_0 = const()[name = string("op_3394_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_3394_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3394_validate_indices_0, x = p2c_att_57_cast_fp16)[name = string("op_3394_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_3394_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3394_validate_indices_0, x = p2c_att_57_cast_fp16)[name = string("op_3394_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_59_perm_0 = const()[name = string("p2c_att_59_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3397_y_0_to_fp16 = const()[name = string("_inversed_3397_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_59_cast_fp16 = transpose(perm = p2c_att_59_perm_0, x = var_3394_cast_fp16_cast_int16)[name = string("transpose_55")];
-            tensor<fp16, [16, 512, 512]> _inversed_3397_cast_fp16 = mul(x = p2c_att_59_cast_fp16, y = _inversed_3397_y_0_to_fp16)[name = string("_inversed_3397_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_29_cast_fp16 = add(x = _inversed_3371_cast_fp16, y = _inversed_3397_cast_fp16)[name = string("rel_att_29_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_87_cast_fp16 = add(x = attention_scores_85_cast_fp16, y = rel_att_29_cast_fp16)[name = string("attention_scores_87_cast_fp16")];
-            tensor<int32, [4]> var_3402 = const()[name = string("op_3402"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_89_cast_fp16 = reshape(shape = var_3402, x = attention_scores_87_cast_fp16)[name = string("attention_scores_89_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_345_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_89_cast_fp16, cond = var_339)[name = string("input_345_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_347_cast_fp16 = softmax(axis = var_56, x = input_345_cast_fp16)[name = string("input_347_cast_fp16")];
-            tensor<int32, [3]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_3412_cast_fp16 = reshape(shape = var_3411, x = input_347_cast_fp16)[name = string("op_3412_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_59_cast_fp16 = transpose(perm = p2c_att_59_perm_0, x = var_3394_cast_fp16_cast_int16)[name = string("transpose_55")];
+            tensor<fp16, [16, 224, 224]> _inversed_3397_cast_fp16 = mul(x = p2c_att_59_cast_fp16, y = _inversed_3397_y_0_to_fp16)[name = string("_inversed_3397_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_29_cast_fp16 = add(x = _inversed_3371_cast_fp16, y = _inversed_3397_cast_fp16)[name = string("rel_att_29_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_87_cast_fp16 = add(x = attention_scores_85_cast_fp16, y = rel_att_29_cast_fp16)[name = string("attention_scores_87_cast_fp16")];
+            tensor<int32, [4]> var_3402 = const()[name = string("op_3402"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_89_cast_fp16 = reshape(shape = var_3402, x = attention_scores_87_cast_fp16)[name = string("attention_scores_89_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_345_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_89_cast_fp16, cond = var_339)[name = string("input_345_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_347_cast_fp16 = softmax(axis = var_56, x = input_345_cast_fp16)[name = string("input_347_cast_fp16")];
+            tensor<int32, [3]> var_3411 = const()[name = string("op_3411"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_3412_cast_fp16 = reshape(shape = var_3411, x = input_347_cast_fp16)[name = string("op_3412_cast_fp16")];
             bool context_layer_57_transpose_x_0 = const()[name = string("context_layer_57_transpose_x_0"), val = bool(false)];
             bool context_layer_57_transpose_y_0 = const()[name = string("context_layer_57_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_57_cast_fp16 = matmul(transpose_x = context_layer_57_transpose_x_0, transpose_y = context_layer_57_transpose_y_0, x = var_3412_cast_fp16, y = value_layer_29_cast_fp16)[name = string("context_layer_57_cast_fp16")];
-            tensor<int32, [4]> var_3416 = const()[name = string("op_3416"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3417_cast_fp16 = reshape(shape = var_3416, x = context_layer_57_cast_fp16)[name = string("op_3417_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_57_cast_fp16 = matmul(transpose_x = context_layer_57_transpose_x_0, transpose_y = context_layer_57_transpose_y_0, x = var_3412_cast_fp16, y = value_layer_29_cast_fp16)[name = string("context_layer_57_cast_fp16")];
+            tensor<int32, [4]> var_3416 = const()[name = string("op_3416"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3417_cast_fp16 = reshape(shape = var_3416, x = context_layer_57_cast_fp16)[name = string("op_3417_cast_fp16")];
             tensor<int32, [4]> var_3418 = const()[name = string("op_3418"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3423 = const()[name = string("op_3423"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_3419_cast_fp16 = transpose(perm = var_3418, x = var_3417_cast_fp16)[name = string("transpose_54")];
-            tensor<fp16, [1, 512, 1024]> input_349_cast_fp16 = reshape(shape = var_3423, x = var_3419_cast_fp16)[name = string("input_349_cast_fp16")];
+            tensor<int32, [3]> var_3423 = const()[name = string("op_3423"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_3419_cast_fp16 = transpose(perm = var_3418, x = var_3417_cast_fp16)[name = string("transpose_54")];
+            tensor<fp16, [1, 224, 1024]> input_349_cast_fp16 = reshape(shape = var_3423, x = var_3419_cast_fp16)[name = string("input_349_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_14_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(653685952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655783168)))];
-            tensor<fp16, [1, 512, 1024]> linear_117_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_attention_output_dense_weight_to_fp16, x = input_349_cast_fp16)[name = string("linear_117_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_353_cast_fp16 = add(x = linear_117_cast_fp16, y = input_341_cast_fp16)[name = string("input_353_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_117_cast_fp16 = linear(bias = encoder_encoder_layer_14_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_attention_output_dense_weight_to_fp16, x = input_349_cast_fp16)[name = string("linear_117_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_353_cast_fp16 = add(x = linear_117_cast_fp16, y = input_341_cast_fp16)[name = string("input_353_cast_fp16")];
             tensor<int32, [1]> input_355_axes_0 = const()[name = string("input_355_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655785280)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655787392)))];
-            tensor<fp16, [1, 512, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, beta = encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_355_cast_fp16 = layer_norm(axes = input_355_axes_0, beta = encoder_encoder_layer_14_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_attention_output_LayerNorm_weight_to_fp16, x = input_353_cast_fp16)[name = string("input_355_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_14_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(655789504)))];
             tensor<fp16, [4096]> encoder_encoder_layer_14_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664178176)))];
-            tensor<fp16, [1, 512, 4096]> linear_118_cast_fp16 = linear(bias = encoder_encoder_layer_14_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_14_intermediate_dense_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_118_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_118_cast_fp16 = linear(bias = encoder_encoder_layer_14_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_14_intermediate_dense_weight_to_fp16, x = input_355_cast_fp16)[name = string("linear_118_cast_fp16")];
             string input_359_mode_0 = const()[name = string("input_359_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_359_cast_fp16 = gelu(mode = input_359_mode_0, x = linear_118_cast_fp16)[name = string("input_359_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_359_cast_fp16 = gelu(mode = input_359_mode_0, x = linear_118_cast_fp16)[name = string("input_359_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_14_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664186432)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672575104)))];
-            tensor<fp16, [1, 512, 1024]> linear_119_cast_fp16 = linear(bias = encoder_encoder_layer_14_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_output_dense_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_119_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_363_cast_fp16 = add(x = linear_119_cast_fp16, y = input_355_cast_fp16)[name = string("input_363_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_119_cast_fp16 = linear(bias = encoder_encoder_layer_14_output_dense_bias_to_fp16, weight = encoder_encoder_layer_14_output_dense_weight_to_fp16, x = input_359_cast_fp16)[name = string("linear_119_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_363_cast_fp16 = add(x = linear_119_cast_fp16, y = input_355_cast_fp16)[name = string("input_363_cast_fp16")];
             tensor<int32, [1]> input_365_axes_0 = const()[name = string("input_365_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672577216)))];
             tensor<fp16, [1024]> encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672579328)))];
-            tensor<fp16, [1, 512, 1024]> input_365_cast_fp16 = layer_norm(axes = input_365_axes_0, beta = encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16, x = input_363_cast_fp16)[name = string("input_365_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_365_cast_fp16 = layer_norm(axes = input_365_axes_0, beta = encoder_encoder_layer_14_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_14_output_LayerNorm_weight_to_fp16, x = input_363_cast_fp16)[name = string("input_365_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(672581440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(674678656)))];
-            tensor<fp16, [1, 512, 1024]> linear_120_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_120_cast_fp16")];
-            tensor<int32, [4]> var_3474 = const()[name = string("op_3474"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_303_cast_fp16 = reshape(shape = var_3474, x = linear_120_cast_fp16)[name = string("x_303_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_120_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_query_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_120_cast_fp16")];
+            tensor<int32, [4]> var_3474 = const()[name = string("op_3474"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_303_cast_fp16 = reshape(shape = var_3474, x = linear_120_cast_fp16)[name = string("x_303_cast_fp16")];
             tensor<int32, [4]> var_3476 = const()[name = string("op_3476"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3477_cast_fp16 = transpose(perm = var_3476, x = x_303_cast_fp16)[name = string("transpose_53")];
-            tensor<fp16, [16, 512, 64]> query_layer_93_cast_fp16 = reshape(shape = var_3481, x = var_3477_cast_fp16)[name = string("query_layer_93_cast_fp16")];
+            tensor<int32, [3]> var_3481 = const()[name = string("op_3481"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3477_cast_fp16 = transpose(perm = var_3476, x = x_303_cast_fp16)[name = string("transpose_53")];
+            tensor<fp16, [16, 224, 64]> query_layer_93_cast_fp16 = reshape(shape = var_3481, x = var_3477_cast_fp16)[name = string("query_layer_93_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(674680768)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676777984)))];
-            tensor<fp16, [1, 512, 1024]> linear_121_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_121_cast_fp16")];
-            tensor<int32, [4]> var_3488 = const()[name = string("op_3488"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_307_cast_fp16 = reshape(shape = var_3488, x = linear_121_cast_fp16)[name = string("x_307_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_121_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_key_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_121_cast_fp16")];
+            tensor<int32, [4]> var_3488 = const()[name = string("op_3488"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_307_cast_fp16 = reshape(shape = var_3488, x = linear_121_cast_fp16)[name = string("x_307_cast_fp16")];
             tensor<int32, [4]> var_3490 = const()[name = string("op_3490"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3495 = const()[name = string("op_3495"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3491_cast_fp16 = transpose(perm = var_3490, x = x_307_cast_fp16)[name = string("transpose_52")];
-            tensor<fp16, [16, 512, 64]> key_layer_31_cast_fp16 = reshape(shape = var_3495, x = var_3491_cast_fp16)[name = string("key_layer_31_cast_fp16")];
+            tensor<int32, [3]> var_3495 = const()[name = string("op_3495"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3491_cast_fp16 = transpose(perm = var_3490, x = x_307_cast_fp16)[name = string("transpose_52")];
+            tensor<fp16, [16, 224, 64]> key_layer_31_cast_fp16 = reshape(shape = var_3495, x = var_3491_cast_fp16)[name = string("key_layer_31_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676780096)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(678877312)))];
-            tensor<fp16, [1, 512, 1024]> linear_122_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_122_cast_fp16")];
-            tensor<int32, [4]> var_3502 = const()[name = string("op_3502"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_311_cast_fp16 = reshape(shape = var_3502, x = linear_122_cast_fp16)[name = string("x_311_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_122_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_15_attention_self_value_proj_weight_to_fp16, x = input_365_cast_fp16)[name = string("linear_122_cast_fp16")];
+            tensor<int32, [4]> var_3502 = const()[name = string("op_3502"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_311_cast_fp16 = reshape(shape = var_3502, x = linear_122_cast_fp16)[name = string("x_311_cast_fp16")];
             tensor<int32, [4]> var_3504 = const()[name = string("op_3504"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3509 = const()[name = string("op_3509"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3505_cast_fp16 = transpose(perm = var_3504, x = x_311_cast_fp16)[name = string("transpose_51")];
-            tensor<fp16, [16, 512, 64]> value_layer_31_cast_fp16 = reshape(shape = var_3509, x = var_3505_cast_fp16)[name = string("value_layer_31_cast_fp16")];
+            tensor<int32, [3]> var_3509 = const()[name = string("op_3509"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3505_cast_fp16 = transpose(perm = var_3504, x = x_311_cast_fp16)[name = string("transpose_51")];
+            tensor<fp16, [16, 224, 64]> value_layer_31_cast_fp16 = reshape(shape = var_3509, x = var_3505_cast_fp16)[name = string("value_layer_31_cast_fp16")];
             tensor<int32, [3]> var_3519_perm_0 = const()[name = string("op_3519_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3521_y_0_to_fp16 = const()[name = string("_inversed_3521_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_3519_cast_fp16 = transpose(perm = var_3519_perm_0, x = key_layer_31_cast_fp16)[name = string("transpose_50")];
-            tensor<fp16, [16, 64, 512]> _inversed_3521_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_3521_y_0_to_fp16)[name = string("_inversed_3521_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_3519_cast_fp16 = transpose(perm = var_3519_perm_0, x = key_layer_31_cast_fp16)[name = string("transpose_50")];
+            tensor<fp16, [16, 64, 224]> _inversed_3521_cast_fp16 = mul(x = var_3519_cast_fp16, y = _inversed_3521_y_0_to_fp16)[name = string("_inversed_3521_cast_fp16")];
             bool attention_scores_91_transpose_x_0 = const()[name = string("attention_scores_91_transpose_x_0"), val = bool(false)];
             bool attention_scores_91_transpose_y_0 = const()[name = string("attention_scores_91_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_91_cast_fp16 = matmul(transpose_x = attention_scores_91_transpose_x_0, transpose_y = attention_scores_91_transpose_y_0, x = query_layer_93_cast_fp16, y = _inversed_3521_cast_fp16)[name = string("attention_scores_91_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_91_cast_fp16 = matmul(transpose_x = attention_scores_91_transpose_x_0, transpose_y = attention_scores_91_transpose_y_0, x = query_layer_93_cast_fp16, y = _inversed_3521_cast_fp16)[name = string("attention_scores_91_cast_fp16")];
             bool c2p_att_61_transpose_x_0 = const()[name = string("c2p_att_61_transpose_x_0"), val = bool(false)];
             bool c2p_att_61_transpose_y_0 = const()[name = string("c2p_att_61_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3577_to_fp16 = const()[name = string("op_3577_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(678879424)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_61_cast_fp16 = matmul(transpose_x = c2p_att_61_transpose_x_0, transpose_y = c2p_att_61_transpose_y_0, x = query_layer_93_cast_fp16, y = var_3577_to_fp16)[name = string("c2p_att_61_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_61_cast_fp16 = matmul(transpose_x = c2p_att_61_transpose_x_0, transpose_y = c2p_att_61_transpose_y_0, x = query_layer_93_cast_fp16, y = var_3577_to_fp16)[name = string("c2p_att_61_cast_fp16")];
             bool c2p_att_63_validate_indices_0 = const()[name = string("c2p_att_63_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_63_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_63_validate_indices_0, x = c2p_att_61_cast_fp16)[name = string("c2p_att_63_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_63_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_63_validate_indices_0, x = c2p_att_61_cast_fp16)[name = string("c2p_att_63_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3590_y_0_to_fp16 = const()[name = string("_inversed_3590_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_3590_cast_fp16 = mul(x = c2p_att_63_cast_fp16_cast_int16, y = _inversed_3590_y_0_to_fp16)[name = string("_inversed_3590_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_3590_cast_fp16 = mul(x = c2p_att_63_cast_fp16_cast_int16, y = _inversed_3590_y_0_to_fp16)[name = string("_inversed_3590_cast_fp16")];
             bool p2c_att_61_transpose_x_0 = const()[name = string("p2c_att_61_transpose_x_0"), val = bool(false)];
             bool p2c_att_61_transpose_y_0 = const()[name = string("p2c_att_61_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3605_to_fp16 = const()[name = string("op_3605_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679928064)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_61_cast_fp16 = matmul(transpose_x = p2c_att_61_transpose_x_0, transpose_y = p2c_att_61_transpose_y_0, x = key_layer_31_cast_fp16, y = var_3605_to_fp16)[name = string("p2c_att_61_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_61_cast_fp16 = matmul(transpose_x = p2c_att_61_transpose_x_0, transpose_y = p2c_att_61_transpose_y_0, x = key_layer_31_cast_fp16, y = var_3605_to_fp16)[name = string("p2c_att_61_cast_fp16")];
             bool var_3613_validate_indices_0 = const()[name = string("op_3613_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_3613_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3613_validate_indices_0, x = p2c_att_61_cast_fp16)[name = string("op_3613_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_3613_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3613_validate_indices_0, x = p2c_att_61_cast_fp16)[name = string("op_3613_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_63_perm_0 = const()[name = string("p2c_att_63_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3616_y_0_to_fp16 = const()[name = string("_inversed_3616_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_63_cast_fp16 = transpose(perm = p2c_att_63_perm_0, x = var_3613_cast_fp16_cast_int16)[name = string("transpose_49")];
-            tensor<fp16, [16, 512, 512]> _inversed_3616_cast_fp16 = mul(x = p2c_att_63_cast_fp16, y = _inversed_3616_y_0_to_fp16)[name = string("_inversed_3616_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_31_cast_fp16 = add(x = _inversed_3590_cast_fp16, y = _inversed_3616_cast_fp16)[name = string("rel_att_31_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_93_cast_fp16 = add(x = attention_scores_91_cast_fp16, y = rel_att_31_cast_fp16)[name = string("attention_scores_93_cast_fp16")];
-            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_95_cast_fp16 = reshape(shape = var_3621, x = attention_scores_93_cast_fp16)[name = string("attention_scores_95_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_369_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_95_cast_fp16, cond = var_339)[name = string("input_369_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_371_cast_fp16 = softmax(axis = var_56, x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
-            tensor<int32, [3]> var_3630 = const()[name = string("op_3630"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_3631_cast_fp16 = reshape(shape = var_3630, x = input_371_cast_fp16)[name = string("op_3631_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_63_cast_fp16 = transpose(perm = p2c_att_63_perm_0, x = var_3613_cast_fp16_cast_int16)[name = string("transpose_49")];
+            tensor<fp16, [16, 224, 224]> _inversed_3616_cast_fp16 = mul(x = p2c_att_63_cast_fp16, y = _inversed_3616_y_0_to_fp16)[name = string("_inversed_3616_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_31_cast_fp16 = add(x = _inversed_3590_cast_fp16, y = _inversed_3616_cast_fp16)[name = string("rel_att_31_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_93_cast_fp16 = add(x = attention_scores_91_cast_fp16, y = rel_att_31_cast_fp16)[name = string("attention_scores_93_cast_fp16")];
+            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_95_cast_fp16 = reshape(shape = var_3621, x = attention_scores_93_cast_fp16)[name = string("attention_scores_95_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_369_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_95_cast_fp16, cond = var_339)[name = string("input_369_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_371_cast_fp16 = softmax(axis = var_56, x = input_369_cast_fp16)[name = string("input_371_cast_fp16")];
+            tensor<int32, [3]> var_3630 = const()[name = string("op_3630"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_3631_cast_fp16 = reshape(shape = var_3630, x = input_371_cast_fp16)[name = string("op_3631_cast_fp16")];
             bool context_layer_61_transpose_x_0 = const()[name = string("context_layer_61_transpose_x_0"), val = bool(false)];
             bool context_layer_61_transpose_y_0 = const()[name = string("context_layer_61_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_61_cast_fp16 = matmul(transpose_x = context_layer_61_transpose_x_0, transpose_y = context_layer_61_transpose_y_0, x = var_3631_cast_fp16, y = value_layer_31_cast_fp16)[name = string("context_layer_61_cast_fp16")];
-            tensor<int32, [4]> var_3635 = const()[name = string("op_3635"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3636_cast_fp16 = reshape(shape = var_3635, x = context_layer_61_cast_fp16)[name = string("op_3636_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_61_cast_fp16 = matmul(transpose_x = context_layer_61_transpose_x_0, transpose_y = context_layer_61_transpose_y_0, x = var_3631_cast_fp16, y = value_layer_31_cast_fp16)[name = string("context_layer_61_cast_fp16")];
+            tensor<int32, [4]> var_3635 = const()[name = string("op_3635"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3636_cast_fp16 = reshape(shape = var_3635, x = context_layer_61_cast_fp16)[name = string("op_3636_cast_fp16")];
             tensor<int32, [4]> var_3637 = const()[name = string("op_3637"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3642 = const()[name = string("op_3642"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_3638_cast_fp16 = transpose(perm = var_3637, x = var_3636_cast_fp16)[name = string("transpose_48")];
-            tensor<fp16, [1, 512, 1024]> input_373_cast_fp16 = reshape(shape = var_3642, x = var_3638_cast_fp16)[name = string("input_373_cast_fp16")];
+            tensor<int32, [3]> var_3642 = const()[name = string("op_3642"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_3638_cast_fp16 = transpose(perm = var_3637, x = var_3636_cast_fp16)[name = string("transpose_48")];
+            tensor<fp16, [1, 224, 1024]> input_373_cast_fp16 = reshape(shape = var_3642, x = var_3638_cast_fp16)[name = string("input_373_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_15_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680976704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683073920)))];
-            tensor<fp16, [1, 512, 1024]> linear_125_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_attention_output_dense_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_125_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_377_cast_fp16 = add(x = linear_125_cast_fp16, y = input_365_cast_fp16)[name = string("input_377_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_125_cast_fp16 = linear(bias = encoder_encoder_layer_15_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_attention_output_dense_weight_to_fp16, x = input_373_cast_fp16)[name = string("linear_125_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_377_cast_fp16 = add(x = linear_125_cast_fp16, y = input_365_cast_fp16)[name = string("input_377_cast_fp16")];
             tensor<int32, [1]> input_379_axes_0 = const()[name = string("input_379_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683076032)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683078144)))];
-            tensor<fp16, [1, 512, 1024]> input_379_cast_fp16 = layer_norm(axes = input_379_axes_0, beta = encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16, x = input_377_cast_fp16)[name = string("input_379_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_379_cast_fp16 = layer_norm(axes = input_379_axes_0, beta = encoder_encoder_layer_15_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_attention_output_LayerNorm_weight_to_fp16, x = input_377_cast_fp16)[name = string("input_379_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_15_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683080256)))];
             tensor<fp16, [4096]> encoder_encoder_layer_15_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691468928)))];
-            tensor<fp16, [1, 512, 4096]> linear_126_cast_fp16 = linear(bias = encoder_encoder_layer_15_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_15_intermediate_dense_weight_to_fp16, x = input_379_cast_fp16)[name = string("linear_126_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_126_cast_fp16 = linear(bias = encoder_encoder_layer_15_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_15_intermediate_dense_weight_to_fp16, x = input_379_cast_fp16)[name = string("linear_126_cast_fp16")];
             string input_383_mode_0 = const()[name = string("input_383_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_383_cast_fp16 = gelu(mode = input_383_mode_0, x = linear_126_cast_fp16)[name = string("input_383_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_383_cast_fp16 = gelu(mode = input_383_mode_0, x = linear_126_cast_fp16)[name = string("input_383_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_15_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691477184)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699865856)))];
-            tensor<fp16, [1, 512, 1024]> linear_127_cast_fp16 = linear(bias = encoder_encoder_layer_15_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_output_dense_weight_to_fp16, x = input_383_cast_fp16)[name = string("linear_127_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_387_cast_fp16 = add(x = linear_127_cast_fp16, y = input_379_cast_fp16)[name = string("input_387_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_127_cast_fp16 = linear(bias = encoder_encoder_layer_15_output_dense_bias_to_fp16, weight = encoder_encoder_layer_15_output_dense_weight_to_fp16, x = input_383_cast_fp16)[name = string("linear_127_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_387_cast_fp16 = add(x = linear_127_cast_fp16, y = input_379_cast_fp16)[name = string("input_387_cast_fp16")];
             tensor<int32, [1]> input_389_axes_0 = const()[name = string("input_389_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699867968)))];
             tensor<fp16, [1024]> encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699870080)))];
-            tensor<fp16, [1, 512, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_389_cast_fp16 = layer_norm(axes = input_389_axes_0, beta = encoder_encoder_layer_15_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_15_output_LayerNorm_weight_to_fp16, x = input_387_cast_fp16)[name = string("input_389_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(699872192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(701969408)))];
-            tensor<fp16, [1, 512, 1024]> linear_128_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_128_cast_fp16")];
-            tensor<int32, [4]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_323_cast_fp16 = reshape(shape = var_3693, x = linear_128_cast_fp16)[name = string("x_323_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_128_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_query_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_128_cast_fp16")];
+            tensor<int32, [4]> var_3693 = const()[name = string("op_3693"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_323_cast_fp16 = reshape(shape = var_3693, x = linear_128_cast_fp16)[name = string("x_323_cast_fp16")];
             tensor<int32, [4]> var_3695 = const()[name = string("op_3695"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3700 = const()[name = string("op_3700"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3696_cast_fp16 = transpose(perm = var_3695, x = x_323_cast_fp16)[name = string("transpose_47")];
-            tensor<fp16, [16, 512, 64]> query_layer_99_cast_fp16 = reshape(shape = var_3700, x = var_3696_cast_fp16)[name = string("query_layer_99_cast_fp16")];
+            tensor<int32, [3]> var_3700 = const()[name = string("op_3700"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3696_cast_fp16 = transpose(perm = var_3695, x = x_323_cast_fp16)[name = string("transpose_47")];
+            tensor<fp16, [16, 224, 64]> query_layer_99_cast_fp16 = reshape(shape = var_3700, x = var_3696_cast_fp16)[name = string("query_layer_99_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(701971520)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704068736)))];
-            tensor<fp16, [1, 512, 1024]> linear_129_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_129_cast_fp16")];
-            tensor<int32, [4]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_327_cast_fp16 = reshape(shape = var_3707, x = linear_129_cast_fp16)[name = string("x_327_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_129_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_key_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_129_cast_fp16")];
+            tensor<int32, [4]> var_3707 = const()[name = string("op_3707"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_327_cast_fp16 = reshape(shape = var_3707, x = linear_129_cast_fp16)[name = string("x_327_cast_fp16")];
             tensor<int32, [4]> var_3709 = const()[name = string("op_3709"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3714 = const()[name = string("op_3714"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3710_cast_fp16 = transpose(perm = var_3709, x = x_327_cast_fp16)[name = string("transpose_46")];
-            tensor<fp16, [16, 512, 64]> key_layer_33_cast_fp16 = reshape(shape = var_3714, x = var_3710_cast_fp16)[name = string("key_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_3714 = const()[name = string("op_3714"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3710_cast_fp16 = transpose(perm = var_3709, x = x_327_cast_fp16)[name = string("transpose_46")];
+            tensor<fp16, [16, 224, 64]> key_layer_33_cast_fp16 = reshape(shape = var_3714, x = var_3710_cast_fp16)[name = string("key_layer_33_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704070848)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706168064)))];
-            tensor<fp16, [1, 512, 1024]> linear_130_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_130_cast_fp16")];
-            tensor<int32, [4]> var_3721 = const()[name = string("op_3721"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_331_cast_fp16 = reshape(shape = var_3721, x = linear_130_cast_fp16)[name = string("x_331_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_130_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_16_attention_self_value_proj_weight_to_fp16, x = input_389_cast_fp16)[name = string("linear_130_cast_fp16")];
+            tensor<int32, [4]> var_3721 = const()[name = string("op_3721"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_331_cast_fp16 = reshape(shape = var_3721, x = linear_130_cast_fp16)[name = string("x_331_cast_fp16")];
             tensor<int32, [4]> var_3723 = const()[name = string("op_3723"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3728 = const()[name = string("op_3728"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3724_cast_fp16 = transpose(perm = var_3723, x = x_331_cast_fp16)[name = string("transpose_45")];
-            tensor<fp16, [16, 512, 64]> value_layer_33_cast_fp16 = reshape(shape = var_3728, x = var_3724_cast_fp16)[name = string("value_layer_33_cast_fp16")];
+            tensor<int32, [3]> var_3728 = const()[name = string("op_3728"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3724_cast_fp16 = transpose(perm = var_3723, x = x_331_cast_fp16)[name = string("transpose_45")];
+            tensor<fp16, [16, 224, 64]> value_layer_33_cast_fp16 = reshape(shape = var_3728, x = var_3724_cast_fp16)[name = string("value_layer_33_cast_fp16")];
             tensor<int32, [3]> var_3738_perm_0 = const()[name = string("op_3738_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3740_y_0_to_fp16 = const()[name = string("_inversed_3740_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_3738_cast_fp16 = transpose(perm = var_3738_perm_0, x = key_layer_33_cast_fp16)[name = string("transpose_44")];
-            tensor<fp16, [16, 64, 512]> _inversed_3740_cast_fp16 = mul(x = var_3738_cast_fp16, y = _inversed_3740_y_0_to_fp16)[name = string("_inversed_3740_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_3738_cast_fp16 = transpose(perm = var_3738_perm_0, x = key_layer_33_cast_fp16)[name = string("transpose_44")];
+            tensor<fp16, [16, 64, 224]> _inversed_3740_cast_fp16 = mul(x = var_3738_cast_fp16, y = _inversed_3740_y_0_to_fp16)[name = string("_inversed_3740_cast_fp16")];
             bool attention_scores_97_transpose_x_0 = const()[name = string("attention_scores_97_transpose_x_0"), val = bool(false)];
             bool attention_scores_97_transpose_y_0 = const()[name = string("attention_scores_97_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_97_cast_fp16 = matmul(transpose_x = attention_scores_97_transpose_x_0, transpose_y = attention_scores_97_transpose_y_0, x = query_layer_99_cast_fp16, y = _inversed_3740_cast_fp16)[name = string("attention_scores_97_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_97_cast_fp16 = matmul(transpose_x = attention_scores_97_transpose_x_0, transpose_y = attention_scores_97_transpose_y_0, x = query_layer_99_cast_fp16, y = _inversed_3740_cast_fp16)[name = string("attention_scores_97_cast_fp16")];
             bool c2p_att_65_transpose_x_0 = const()[name = string("c2p_att_65_transpose_x_0"), val = bool(false)];
             bool c2p_att_65_transpose_y_0 = const()[name = string("c2p_att_65_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3796_to_fp16 = const()[name = string("op_3796_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(706170176)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_65_cast_fp16 = matmul(transpose_x = c2p_att_65_transpose_x_0, transpose_y = c2p_att_65_transpose_y_0, x = query_layer_99_cast_fp16, y = var_3796_to_fp16)[name = string("c2p_att_65_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_65_cast_fp16 = matmul(transpose_x = c2p_att_65_transpose_x_0, transpose_y = c2p_att_65_transpose_y_0, x = query_layer_99_cast_fp16, y = var_3796_to_fp16)[name = string("c2p_att_65_cast_fp16")];
             bool c2p_att_67_validate_indices_0 = const()[name = string("c2p_att_67_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_67_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_67_validate_indices_0, x = c2p_att_65_cast_fp16)[name = string("c2p_att_67_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_67_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_67_validate_indices_0, x = c2p_att_65_cast_fp16)[name = string("c2p_att_67_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_3809_y_0_to_fp16 = const()[name = string("_inversed_3809_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_3809_cast_fp16 = mul(x = c2p_att_67_cast_fp16_cast_int16, y = _inversed_3809_y_0_to_fp16)[name = string("_inversed_3809_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_3809_cast_fp16 = mul(x = c2p_att_67_cast_fp16_cast_int16, y = _inversed_3809_y_0_to_fp16)[name = string("_inversed_3809_cast_fp16")];
             bool p2c_att_65_transpose_x_0 = const()[name = string("p2c_att_65_transpose_x_0"), val = bool(false)];
             bool p2c_att_65_transpose_y_0 = const()[name = string("p2c_att_65_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_3824_to_fp16 = const()[name = string("op_3824_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707218816)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_65_cast_fp16 = matmul(transpose_x = p2c_att_65_transpose_x_0, transpose_y = p2c_att_65_transpose_y_0, x = key_layer_33_cast_fp16, y = var_3824_to_fp16)[name = string("p2c_att_65_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_65_cast_fp16 = matmul(transpose_x = p2c_att_65_transpose_x_0, transpose_y = p2c_att_65_transpose_y_0, x = key_layer_33_cast_fp16, y = var_3824_to_fp16)[name = string("p2c_att_65_cast_fp16")];
             bool var_3832_validate_indices_0 = const()[name = string("op_3832_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_3832_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3832_validate_indices_0, x = p2c_att_65_cast_fp16)[name = string("op_3832_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_3832_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_3832_validate_indices_0, x = p2c_att_65_cast_fp16)[name = string("op_3832_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_67_perm_0 = const()[name = string("p2c_att_67_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3835_y_0_to_fp16 = const()[name = string("_inversed_3835_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_67_cast_fp16 = transpose(perm = p2c_att_67_perm_0, x = var_3832_cast_fp16_cast_int16)[name = string("transpose_43")];
-            tensor<fp16, [16, 512, 512]> _inversed_3835_cast_fp16 = mul(x = p2c_att_67_cast_fp16, y = _inversed_3835_y_0_to_fp16)[name = string("_inversed_3835_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_33_cast_fp16 = add(x = _inversed_3809_cast_fp16, y = _inversed_3835_cast_fp16)[name = string("rel_att_33_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_99_cast_fp16 = add(x = attention_scores_97_cast_fp16, y = rel_att_33_cast_fp16)[name = string("attention_scores_99_cast_fp16")];
-            tensor<int32, [4]> var_3840 = const()[name = string("op_3840"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_101_cast_fp16 = reshape(shape = var_3840, x = attention_scores_99_cast_fp16)[name = string("attention_scores_101_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_393_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_101_cast_fp16, cond = var_339)[name = string("input_393_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_395_cast_fp16 = softmax(axis = var_56, x = input_393_cast_fp16)[name = string("input_395_cast_fp16")];
-            tensor<int32, [3]> var_3849 = const()[name = string("op_3849"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_3850_cast_fp16 = reshape(shape = var_3849, x = input_395_cast_fp16)[name = string("op_3850_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_67_cast_fp16 = transpose(perm = p2c_att_67_perm_0, x = var_3832_cast_fp16_cast_int16)[name = string("transpose_43")];
+            tensor<fp16, [16, 224, 224]> _inversed_3835_cast_fp16 = mul(x = p2c_att_67_cast_fp16, y = _inversed_3835_y_0_to_fp16)[name = string("_inversed_3835_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_33_cast_fp16 = add(x = _inversed_3809_cast_fp16, y = _inversed_3835_cast_fp16)[name = string("rel_att_33_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_99_cast_fp16 = add(x = attention_scores_97_cast_fp16, y = rel_att_33_cast_fp16)[name = string("attention_scores_99_cast_fp16")];
+            tensor<int32, [4]> var_3840 = const()[name = string("op_3840"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_101_cast_fp16 = reshape(shape = var_3840, x = attention_scores_99_cast_fp16)[name = string("attention_scores_101_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_393_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_101_cast_fp16, cond = var_339)[name = string("input_393_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_395_cast_fp16 = softmax(axis = var_56, x = input_393_cast_fp16)[name = string("input_395_cast_fp16")];
+            tensor<int32, [3]> var_3849 = const()[name = string("op_3849"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_3850_cast_fp16 = reshape(shape = var_3849, x = input_395_cast_fp16)[name = string("op_3850_cast_fp16")];
             bool context_layer_65_transpose_x_0 = const()[name = string("context_layer_65_transpose_x_0"), val = bool(false)];
             bool context_layer_65_transpose_y_0 = const()[name = string("context_layer_65_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_65_cast_fp16 = matmul(transpose_x = context_layer_65_transpose_x_0, transpose_y = context_layer_65_transpose_y_0, x = var_3850_cast_fp16, y = value_layer_33_cast_fp16)[name = string("context_layer_65_cast_fp16")];
-            tensor<int32, [4]> var_3854 = const()[name = string("op_3854"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3855_cast_fp16 = reshape(shape = var_3854, x = context_layer_65_cast_fp16)[name = string("op_3855_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_65_cast_fp16 = matmul(transpose_x = context_layer_65_transpose_x_0, transpose_y = context_layer_65_transpose_y_0, x = var_3850_cast_fp16, y = value_layer_33_cast_fp16)[name = string("context_layer_65_cast_fp16")];
+            tensor<int32, [4]> var_3854 = const()[name = string("op_3854"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3855_cast_fp16 = reshape(shape = var_3854, x = context_layer_65_cast_fp16)[name = string("op_3855_cast_fp16")];
             tensor<int32, [4]> var_3856 = const()[name = string("op_3856"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_3857_cast_fp16 = transpose(perm = var_3856, x = var_3855_cast_fp16)[name = string("transpose_42")];
-            tensor<fp16, [1, 512, 1024]> input_397_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_397_cast_fp16")];
+            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_3857_cast_fp16 = transpose(perm = var_3856, x = var_3855_cast_fp16)[name = string("transpose_42")];
+            tensor<fp16, [1, 224, 1024]> input_397_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_397_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_16_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708267456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710364672)))];
-            tensor<fp16, [1, 512, 1024]> linear_133_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_attention_output_dense_weight_to_fp16, x = input_397_cast_fp16)[name = string("linear_133_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_401_cast_fp16 = add(x = linear_133_cast_fp16, y = input_389_cast_fp16)[name = string("input_401_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_133_cast_fp16 = linear(bias = encoder_encoder_layer_16_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_attention_output_dense_weight_to_fp16, x = input_397_cast_fp16)[name = string("linear_133_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_401_cast_fp16 = add(x = linear_133_cast_fp16, y = input_389_cast_fp16)[name = string("input_401_cast_fp16")];
             tensor<int32, [1]> input_403_axes_0 = const()[name = string("input_403_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710366784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710368896)))];
-            tensor<fp16, [1, 512, 1024]> input_403_cast_fp16 = layer_norm(axes = input_403_axes_0, beta = encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16, x = input_401_cast_fp16)[name = string("input_403_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_403_cast_fp16 = layer_norm(axes = input_403_axes_0, beta = encoder_encoder_layer_16_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_attention_output_LayerNorm_weight_to_fp16, x = input_401_cast_fp16)[name = string("input_403_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_16_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(710371008)))];
             tensor<fp16, [4096]> encoder_encoder_layer_16_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(718759680)))];
-            tensor<fp16, [1, 512, 4096]> linear_134_cast_fp16 = linear(bias = encoder_encoder_layer_16_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_16_intermediate_dense_weight_to_fp16, x = input_403_cast_fp16)[name = string("linear_134_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_134_cast_fp16 = linear(bias = encoder_encoder_layer_16_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_16_intermediate_dense_weight_to_fp16, x = input_403_cast_fp16)[name = string("linear_134_cast_fp16")];
             string input_407_mode_0 = const()[name = string("input_407_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_407_cast_fp16 = gelu(mode = input_407_mode_0, x = linear_134_cast_fp16)[name = string("input_407_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_407_cast_fp16 = gelu(mode = input_407_mode_0, x = linear_134_cast_fp16)[name = string("input_407_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_16_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(718767936)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727156608)))];
-            tensor<fp16, [1, 512, 1024]> linear_135_cast_fp16 = linear(bias = encoder_encoder_layer_16_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_output_dense_weight_to_fp16, x = input_407_cast_fp16)[name = string("linear_135_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_411_cast_fp16 = add(x = linear_135_cast_fp16, y = input_403_cast_fp16)[name = string("input_411_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_135_cast_fp16 = linear(bias = encoder_encoder_layer_16_output_dense_bias_to_fp16, weight = encoder_encoder_layer_16_output_dense_weight_to_fp16, x = input_407_cast_fp16)[name = string("linear_135_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_411_cast_fp16 = add(x = linear_135_cast_fp16, y = input_403_cast_fp16)[name = string("input_411_cast_fp16")];
             tensor<int32, [1]> input_413_axes_0 = const()[name = string("input_413_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727158720)))];
             tensor<fp16, [1024]> encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727160832)))];
-            tensor<fp16, [1, 512, 1024]> input_413_cast_fp16 = layer_norm(axes = input_413_axes_0, beta = encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16, x = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_413_cast_fp16 = layer_norm(axes = input_413_axes_0, beta = encoder_encoder_layer_16_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_16_output_LayerNorm_weight_to_fp16, x = input_411_cast_fp16)[name = string("input_413_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727162944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729260160)))];
-            tensor<fp16, [1, 512, 1024]> linear_136_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_136_cast_fp16")];
-            tensor<int32, [4]> var_3912 = const()[name = string("op_3912"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_343_cast_fp16 = reshape(shape = var_3912, x = linear_136_cast_fp16)[name = string("x_343_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_136_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_query_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_136_cast_fp16")];
+            tensor<int32, [4]> var_3912 = const()[name = string("op_3912"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_343_cast_fp16 = reshape(shape = var_3912, x = linear_136_cast_fp16)[name = string("x_343_cast_fp16")];
             tensor<int32, [4]> var_3914 = const()[name = string("op_3914"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3919 = const()[name = string("op_3919"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3915_cast_fp16 = transpose(perm = var_3914, x = x_343_cast_fp16)[name = string("transpose_41")];
-            tensor<fp16, [16, 512, 64]> query_layer_105_cast_fp16 = reshape(shape = var_3919, x = var_3915_cast_fp16)[name = string("query_layer_105_cast_fp16")];
+            tensor<int32, [3]> var_3919 = const()[name = string("op_3919"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3915_cast_fp16 = transpose(perm = var_3914, x = x_343_cast_fp16)[name = string("transpose_41")];
+            tensor<fp16, [16, 224, 64]> query_layer_105_cast_fp16 = reshape(shape = var_3919, x = var_3915_cast_fp16)[name = string("query_layer_105_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729262272)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731359488)))];
-            tensor<fp16, [1, 512, 1024]> linear_137_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_137_cast_fp16")];
-            tensor<int32, [4]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_347_cast_fp16 = reshape(shape = var_3926, x = linear_137_cast_fp16)[name = string("x_347_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_137_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_key_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_137_cast_fp16")];
+            tensor<int32, [4]> var_3926 = const()[name = string("op_3926"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_347_cast_fp16 = reshape(shape = var_3926, x = linear_137_cast_fp16)[name = string("x_347_cast_fp16")];
             tensor<int32, [4]> var_3928 = const()[name = string("op_3928"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3929_cast_fp16 = transpose(perm = var_3928, x = x_347_cast_fp16)[name = string("transpose_40")];
-            tensor<fp16, [16, 512, 64]> key_layer_35_cast_fp16 = reshape(shape = var_3933, x = var_3929_cast_fp16)[name = string("key_layer_35_cast_fp16")];
+            tensor<int32, [3]> var_3933 = const()[name = string("op_3933"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3929_cast_fp16 = transpose(perm = var_3928, x = x_347_cast_fp16)[name = string("transpose_40")];
+            tensor<fp16, [16, 224, 64]> key_layer_35_cast_fp16 = reshape(shape = var_3933, x = var_3929_cast_fp16)[name = string("key_layer_35_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(731361600)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733458816)))];
-            tensor<fp16, [1, 512, 1024]> linear_138_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_138_cast_fp16")];
-            tensor<int32, [4]> var_3940 = const()[name = string("op_3940"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_351_cast_fp16 = reshape(shape = var_3940, x = linear_138_cast_fp16)[name = string("x_351_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_138_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_17_attention_self_value_proj_weight_to_fp16, x = input_413_cast_fp16)[name = string("linear_138_cast_fp16")];
+            tensor<int32, [4]> var_3940 = const()[name = string("op_3940"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_351_cast_fp16 = reshape(shape = var_3940, x = linear_138_cast_fp16)[name = string("x_351_cast_fp16")];
             tensor<int32, [4]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_3947 = const()[name = string("op_3947"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_3943_cast_fp16 = transpose(perm = var_3942, x = x_351_cast_fp16)[name = string("transpose_39")];
-            tensor<fp16, [16, 512, 64]> value_layer_35_cast_fp16 = reshape(shape = var_3947, x = var_3943_cast_fp16)[name = string("value_layer_35_cast_fp16")];
+            tensor<int32, [3]> var_3947 = const()[name = string("op_3947"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_3943_cast_fp16 = transpose(perm = var_3942, x = x_351_cast_fp16)[name = string("transpose_39")];
+            tensor<fp16, [16, 224, 64]> value_layer_35_cast_fp16 = reshape(shape = var_3947, x = var_3943_cast_fp16)[name = string("value_layer_35_cast_fp16")];
             tensor<int32, [3]> var_3957_perm_0 = const()[name = string("op_3957_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_3959_y_0_to_fp16 = const()[name = string("_inversed_3959_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = key_layer_35_cast_fp16)[name = string("transpose_38")];
-            tensor<fp16, [16, 64, 512]> _inversed_3959_cast_fp16 = mul(x = var_3957_cast_fp16, y = _inversed_3959_y_0_to_fp16)[name = string("_inversed_3959_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_3957_cast_fp16 = transpose(perm = var_3957_perm_0, x = key_layer_35_cast_fp16)[name = string("transpose_38")];
+            tensor<fp16, [16, 64, 224]> _inversed_3959_cast_fp16 = mul(x = var_3957_cast_fp16, y = _inversed_3959_y_0_to_fp16)[name = string("_inversed_3959_cast_fp16")];
             bool attention_scores_103_transpose_x_0 = const()[name = string("attention_scores_103_transpose_x_0"), val = bool(false)];
             bool attention_scores_103_transpose_y_0 = const()[name = string("attention_scores_103_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_103_cast_fp16 = matmul(transpose_x = attention_scores_103_transpose_x_0, transpose_y = attention_scores_103_transpose_y_0, x = query_layer_105_cast_fp16, y = _inversed_3959_cast_fp16)[name = string("attention_scores_103_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_103_cast_fp16 = matmul(transpose_x = attention_scores_103_transpose_x_0, transpose_y = attention_scores_103_transpose_y_0, x = query_layer_105_cast_fp16, y = _inversed_3959_cast_fp16)[name = string("attention_scores_103_cast_fp16")];
             bool c2p_att_69_transpose_x_0 = const()[name = string("c2p_att_69_transpose_x_0"), val = bool(false)];
             bool c2p_att_69_transpose_y_0 = const()[name = string("c2p_att_69_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4015_to_fp16 = const()[name = string("op_4015_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733460928)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_69_cast_fp16 = matmul(transpose_x = c2p_att_69_transpose_x_0, transpose_y = c2p_att_69_transpose_y_0, x = query_layer_105_cast_fp16, y = var_4015_to_fp16)[name = string("c2p_att_69_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_69_cast_fp16 = matmul(transpose_x = c2p_att_69_transpose_x_0, transpose_y = c2p_att_69_transpose_y_0, x = query_layer_105_cast_fp16, y = var_4015_to_fp16)[name = string("c2p_att_69_cast_fp16")];
             bool c2p_att_71_validate_indices_0 = const()[name = string("c2p_att_71_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_71_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_71_validate_indices_0, x = c2p_att_69_cast_fp16)[name = string("c2p_att_71_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_71_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_71_validate_indices_0, x = c2p_att_69_cast_fp16)[name = string("c2p_att_71_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4028_y_0_to_fp16 = const()[name = string("_inversed_4028_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_4028_cast_fp16 = mul(x = c2p_att_71_cast_fp16_cast_int16, y = _inversed_4028_y_0_to_fp16)[name = string("_inversed_4028_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_4028_cast_fp16 = mul(x = c2p_att_71_cast_fp16_cast_int16, y = _inversed_4028_y_0_to_fp16)[name = string("_inversed_4028_cast_fp16")];
             bool p2c_att_69_transpose_x_0 = const()[name = string("p2c_att_69_transpose_x_0"), val = bool(false)];
             bool p2c_att_69_transpose_y_0 = const()[name = string("p2c_att_69_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4043_to_fp16 = const()[name = string("op_4043_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734509568)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_69_cast_fp16 = matmul(transpose_x = p2c_att_69_transpose_x_0, transpose_y = p2c_att_69_transpose_y_0, x = key_layer_35_cast_fp16, y = var_4043_to_fp16)[name = string("p2c_att_69_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_69_cast_fp16 = matmul(transpose_x = p2c_att_69_transpose_x_0, transpose_y = p2c_att_69_transpose_y_0, x = key_layer_35_cast_fp16, y = var_4043_to_fp16)[name = string("p2c_att_69_cast_fp16")];
             bool var_4051_validate_indices_0 = const()[name = string("op_4051_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_4051_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4051_validate_indices_0, x = p2c_att_69_cast_fp16)[name = string("op_4051_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_4051_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4051_validate_indices_0, x = p2c_att_69_cast_fp16)[name = string("op_4051_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_71_perm_0 = const()[name = string("p2c_att_71_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4054_y_0_to_fp16 = const()[name = string("_inversed_4054_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_71_cast_fp16 = transpose(perm = p2c_att_71_perm_0, x = var_4051_cast_fp16_cast_int16)[name = string("transpose_37")];
-            tensor<fp16, [16, 512, 512]> _inversed_4054_cast_fp16 = mul(x = p2c_att_71_cast_fp16, y = _inversed_4054_y_0_to_fp16)[name = string("_inversed_4054_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_35_cast_fp16 = add(x = _inversed_4028_cast_fp16, y = _inversed_4054_cast_fp16)[name = string("rel_att_35_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_105_cast_fp16 = add(x = attention_scores_103_cast_fp16, y = rel_att_35_cast_fp16)[name = string("attention_scores_105_cast_fp16")];
-            tensor<int32, [4]> var_4059 = const()[name = string("op_4059"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_107_cast_fp16 = reshape(shape = var_4059, x = attention_scores_105_cast_fp16)[name = string("attention_scores_107_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_417_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_107_cast_fp16, cond = var_339)[name = string("input_417_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_419_cast_fp16 = softmax(axis = var_56, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
-            tensor<int32, [3]> var_4068 = const()[name = string("op_4068"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_4069_cast_fp16 = reshape(shape = var_4068, x = input_419_cast_fp16)[name = string("op_4069_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_71_cast_fp16 = transpose(perm = p2c_att_71_perm_0, x = var_4051_cast_fp16_cast_int16)[name = string("transpose_37")];
+            tensor<fp16, [16, 224, 224]> _inversed_4054_cast_fp16 = mul(x = p2c_att_71_cast_fp16, y = _inversed_4054_y_0_to_fp16)[name = string("_inversed_4054_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_35_cast_fp16 = add(x = _inversed_4028_cast_fp16, y = _inversed_4054_cast_fp16)[name = string("rel_att_35_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_105_cast_fp16 = add(x = attention_scores_103_cast_fp16, y = rel_att_35_cast_fp16)[name = string("attention_scores_105_cast_fp16")];
+            tensor<int32, [4]> var_4059 = const()[name = string("op_4059"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_107_cast_fp16 = reshape(shape = var_4059, x = attention_scores_105_cast_fp16)[name = string("attention_scores_107_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_417_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_107_cast_fp16, cond = var_339)[name = string("input_417_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_419_cast_fp16 = softmax(axis = var_56, x = input_417_cast_fp16)[name = string("input_419_cast_fp16")];
+            tensor<int32, [3]> var_4068 = const()[name = string("op_4068"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_4069_cast_fp16 = reshape(shape = var_4068, x = input_419_cast_fp16)[name = string("op_4069_cast_fp16")];
             bool context_layer_69_transpose_x_0 = const()[name = string("context_layer_69_transpose_x_0"), val = bool(false)];
             bool context_layer_69_transpose_y_0 = const()[name = string("context_layer_69_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_69_cast_fp16 = matmul(transpose_x = context_layer_69_transpose_x_0, transpose_y = context_layer_69_transpose_y_0, x = var_4069_cast_fp16, y = value_layer_35_cast_fp16)[name = string("context_layer_69_cast_fp16")];
-            tensor<int32, [4]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4074_cast_fp16 = reshape(shape = var_4073, x = context_layer_69_cast_fp16)[name = string("op_4074_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_69_cast_fp16 = matmul(transpose_x = context_layer_69_transpose_x_0, transpose_y = context_layer_69_transpose_y_0, x = var_4069_cast_fp16, y = value_layer_35_cast_fp16)[name = string("context_layer_69_cast_fp16")];
+            tensor<int32, [4]> var_4073 = const()[name = string("op_4073"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4074_cast_fp16 = reshape(shape = var_4073, x = context_layer_69_cast_fp16)[name = string("op_4074_cast_fp16")];
             tensor<int32, [4]> var_4075 = const()[name = string("op_4075"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_4076_cast_fp16 = transpose(perm = var_4075, x = var_4074_cast_fp16)[name = string("transpose_36")];
-            tensor<fp16, [1, 512, 1024]> input_421_cast_fp16 = reshape(shape = var_4080, x = var_4076_cast_fp16)[name = string("input_421_cast_fp16")];
+            tensor<int32, [3]> var_4080 = const()[name = string("op_4080"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_4076_cast_fp16 = transpose(perm = var_4075, x = var_4074_cast_fp16)[name = string("transpose_36")];
+            tensor<fp16, [1, 224, 1024]> input_421_cast_fp16 = reshape(shape = var_4080, x = var_4076_cast_fp16)[name = string("input_421_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_17_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(735558208)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737655424)))];
-            tensor<fp16, [1, 512, 1024]> linear_141_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_attention_output_dense_weight_to_fp16, x = input_421_cast_fp16)[name = string("linear_141_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_425_cast_fp16 = add(x = linear_141_cast_fp16, y = input_413_cast_fp16)[name = string("input_425_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_141_cast_fp16 = linear(bias = encoder_encoder_layer_17_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_attention_output_dense_weight_to_fp16, x = input_421_cast_fp16)[name = string("linear_141_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_425_cast_fp16 = add(x = linear_141_cast_fp16, y = input_413_cast_fp16)[name = string("input_425_cast_fp16")];
             tensor<int32, [1]> input_427_axes_0 = const()[name = string("input_427_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737657536)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737659648)))];
-            tensor<fp16, [1, 512, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, beta = encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_427_cast_fp16 = layer_norm(axes = input_427_axes_0, beta = encoder_encoder_layer_17_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_attention_output_LayerNorm_weight_to_fp16, x = input_425_cast_fp16)[name = string("input_427_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_17_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(737661760)))];
             tensor<fp16, [4096]> encoder_encoder_layer_17_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746050432)))];
-            tensor<fp16, [1, 512, 4096]> linear_142_cast_fp16 = linear(bias = encoder_encoder_layer_17_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_17_intermediate_dense_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_142_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_142_cast_fp16 = linear(bias = encoder_encoder_layer_17_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_17_intermediate_dense_weight_to_fp16, x = input_427_cast_fp16)[name = string("linear_142_cast_fp16")];
             string input_431_mode_0 = const()[name = string("input_431_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_431_cast_fp16 = gelu(mode = input_431_mode_0, x = linear_142_cast_fp16)[name = string("input_431_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_431_cast_fp16 = gelu(mode = input_431_mode_0, x = linear_142_cast_fp16)[name = string("input_431_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_17_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746058688)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754447360)))];
-            tensor<fp16, [1, 512, 1024]> linear_143_cast_fp16 = linear(bias = encoder_encoder_layer_17_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_output_dense_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_143_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_435_cast_fp16 = add(x = linear_143_cast_fp16, y = input_427_cast_fp16)[name = string("input_435_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_143_cast_fp16 = linear(bias = encoder_encoder_layer_17_output_dense_bias_to_fp16, weight = encoder_encoder_layer_17_output_dense_weight_to_fp16, x = input_431_cast_fp16)[name = string("linear_143_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_435_cast_fp16 = add(x = linear_143_cast_fp16, y = input_427_cast_fp16)[name = string("input_435_cast_fp16")];
             tensor<int32, [1]> input_437_axes_0 = const()[name = string("input_437_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754449472)))];
             tensor<fp16, [1024]> encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754451584)))];
-            tensor<fp16, [1, 512, 1024]> input_437_cast_fp16 = layer_norm(axes = input_437_axes_0, beta = encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16, x = input_435_cast_fp16)[name = string("input_437_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_437_cast_fp16 = layer_norm(axes = input_437_axes_0, beta = encoder_encoder_layer_17_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_17_output_LayerNorm_weight_to_fp16, x = input_435_cast_fp16)[name = string("input_437_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(754453696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(756550912)))];
-            tensor<fp16, [1, 512, 1024]> linear_144_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_144_cast_fp16")];
-            tensor<int32, [4]> var_4131 = const()[name = string("op_4131"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_363_cast_fp16 = reshape(shape = var_4131, x = linear_144_cast_fp16)[name = string("x_363_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_144_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_query_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_144_cast_fp16")];
+            tensor<int32, [4]> var_4131 = const()[name = string("op_4131"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_363_cast_fp16 = reshape(shape = var_4131, x = linear_144_cast_fp16)[name = string("x_363_cast_fp16")];
             tensor<int32, [4]> var_4133 = const()[name = string("op_4133"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4138 = const()[name = string("op_4138"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4134_cast_fp16 = transpose(perm = var_4133, x = x_363_cast_fp16)[name = string("transpose_35")];
-            tensor<fp16, [16, 512, 64]> query_layer_111_cast_fp16 = reshape(shape = var_4138, x = var_4134_cast_fp16)[name = string("query_layer_111_cast_fp16")];
+            tensor<int32, [3]> var_4138 = const()[name = string("op_4138"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4134_cast_fp16 = transpose(perm = var_4133, x = x_363_cast_fp16)[name = string("transpose_35")];
+            tensor<fp16, [16, 224, 64]> query_layer_111_cast_fp16 = reshape(shape = var_4138, x = var_4134_cast_fp16)[name = string("query_layer_111_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(756553024)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758650240)))];
-            tensor<fp16, [1, 512, 1024]> linear_145_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_145_cast_fp16")];
-            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_367_cast_fp16 = reshape(shape = var_4145, x = linear_145_cast_fp16)[name = string("x_367_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_145_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_key_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_145_cast_fp16")];
+            tensor<int32, [4]> var_4145 = const()[name = string("op_4145"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_367_cast_fp16 = reshape(shape = var_4145, x = linear_145_cast_fp16)[name = string("x_367_cast_fp16")];
             tensor<int32, [4]> var_4147 = const()[name = string("op_4147"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4152 = const()[name = string("op_4152"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4148_cast_fp16 = transpose(perm = var_4147, x = x_367_cast_fp16)[name = string("transpose_34")];
-            tensor<fp16, [16, 512, 64]> key_layer_37_cast_fp16 = reshape(shape = var_4152, x = var_4148_cast_fp16)[name = string("key_layer_37_cast_fp16")];
+            tensor<int32, [3]> var_4152 = const()[name = string("op_4152"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4148_cast_fp16 = transpose(perm = var_4147, x = x_367_cast_fp16)[name = string("transpose_34")];
+            tensor<fp16, [16, 224, 64]> key_layer_37_cast_fp16 = reshape(shape = var_4152, x = var_4148_cast_fp16)[name = string("key_layer_37_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(758652352)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(760749568)))];
-            tensor<fp16, [1, 512, 1024]> linear_146_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_146_cast_fp16")];
-            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_371_cast_fp16 = reshape(shape = var_4159, x = linear_146_cast_fp16)[name = string("x_371_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_146_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_18_attention_self_value_proj_weight_to_fp16, x = input_437_cast_fp16)[name = string("linear_146_cast_fp16")];
+            tensor<int32, [4]> var_4159 = const()[name = string("op_4159"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_371_cast_fp16 = reshape(shape = var_4159, x = linear_146_cast_fp16)[name = string("x_371_cast_fp16")];
             tensor<int32, [4]> var_4161 = const()[name = string("op_4161"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4166 = const()[name = string("op_4166"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4162_cast_fp16 = transpose(perm = var_4161, x = x_371_cast_fp16)[name = string("transpose_33")];
-            tensor<fp16, [16, 512, 64]> value_layer_37_cast_fp16 = reshape(shape = var_4166, x = var_4162_cast_fp16)[name = string("value_layer_37_cast_fp16")];
+            tensor<int32, [3]> var_4166 = const()[name = string("op_4166"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4162_cast_fp16 = transpose(perm = var_4161, x = x_371_cast_fp16)[name = string("transpose_33")];
+            tensor<fp16, [16, 224, 64]> value_layer_37_cast_fp16 = reshape(shape = var_4166, x = var_4162_cast_fp16)[name = string("value_layer_37_cast_fp16")];
             tensor<int32, [3]> var_4176_perm_0 = const()[name = string("op_4176_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4178_y_0_to_fp16 = const()[name = string("_inversed_4178_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_4176_cast_fp16 = transpose(perm = var_4176_perm_0, x = key_layer_37_cast_fp16)[name = string("transpose_32")];
-            tensor<fp16, [16, 64, 512]> _inversed_4178_cast_fp16 = mul(x = var_4176_cast_fp16, y = _inversed_4178_y_0_to_fp16)[name = string("_inversed_4178_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_4176_cast_fp16 = transpose(perm = var_4176_perm_0, x = key_layer_37_cast_fp16)[name = string("transpose_32")];
+            tensor<fp16, [16, 64, 224]> _inversed_4178_cast_fp16 = mul(x = var_4176_cast_fp16, y = _inversed_4178_y_0_to_fp16)[name = string("_inversed_4178_cast_fp16")];
             bool attention_scores_109_transpose_x_0 = const()[name = string("attention_scores_109_transpose_x_0"), val = bool(false)];
             bool attention_scores_109_transpose_y_0 = const()[name = string("attention_scores_109_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_109_cast_fp16 = matmul(transpose_x = attention_scores_109_transpose_x_0, transpose_y = attention_scores_109_transpose_y_0, x = query_layer_111_cast_fp16, y = _inversed_4178_cast_fp16)[name = string("attention_scores_109_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_109_cast_fp16 = matmul(transpose_x = attention_scores_109_transpose_x_0, transpose_y = attention_scores_109_transpose_y_0, x = query_layer_111_cast_fp16, y = _inversed_4178_cast_fp16)[name = string("attention_scores_109_cast_fp16")];
             bool c2p_att_73_transpose_x_0 = const()[name = string("c2p_att_73_transpose_x_0"), val = bool(false)];
             bool c2p_att_73_transpose_y_0 = const()[name = string("c2p_att_73_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4234_to_fp16 = const()[name = string("op_4234_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(760751680)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_73_cast_fp16 = matmul(transpose_x = c2p_att_73_transpose_x_0, transpose_y = c2p_att_73_transpose_y_0, x = query_layer_111_cast_fp16, y = var_4234_to_fp16)[name = string("c2p_att_73_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_73_cast_fp16 = matmul(transpose_x = c2p_att_73_transpose_x_0, transpose_y = c2p_att_73_transpose_y_0, x = query_layer_111_cast_fp16, y = var_4234_to_fp16)[name = string("c2p_att_73_cast_fp16")];
             bool c2p_att_75_validate_indices_0 = const()[name = string("c2p_att_75_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_75_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_75_validate_indices_0, x = c2p_att_73_cast_fp16)[name = string("c2p_att_75_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_75_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_75_validate_indices_0, x = c2p_att_73_cast_fp16)[name = string("c2p_att_75_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4247_y_0_to_fp16 = const()[name = string("_inversed_4247_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_4247_cast_fp16 = mul(x = c2p_att_75_cast_fp16_cast_int16, y = _inversed_4247_y_0_to_fp16)[name = string("_inversed_4247_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_4247_cast_fp16 = mul(x = c2p_att_75_cast_fp16_cast_int16, y = _inversed_4247_y_0_to_fp16)[name = string("_inversed_4247_cast_fp16")];
             bool p2c_att_73_transpose_x_0 = const()[name = string("p2c_att_73_transpose_x_0"), val = bool(false)];
             bool p2c_att_73_transpose_y_0 = const()[name = string("p2c_att_73_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4262_to_fp16 = const()[name = string("op_4262_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(761800320)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_73_cast_fp16 = matmul(transpose_x = p2c_att_73_transpose_x_0, transpose_y = p2c_att_73_transpose_y_0, x = key_layer_37_cast_fp16, y = var_4262_to_fp16)[name = string("p2c_att_73_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_73_cast_fp16 = matmul(transpose_x = p2c_att_73_transpose_x_0, transpose_y = p2c_att_73_transpose_y_0, x = key_layer_37_cast_fp16, y = var_4262_to_fp16)[name = string("p2c_att_73_cast_fp16")];
             bool var_4270_validate_indices_0 = const()[name = string("op_4270_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_4270_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4270_validate_indices_0, x = p2c_att_73_cast_fp16)[name = string("op_4270_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_4270_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4270_validate_indices_0, x = p2c_att_73_cast_fp16)[name = string("op_4270_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_75_perm_0 = const()[name = string("p2c_att_75_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4273_y_0_to_fp16 = const()[name = string("_inversed_4273_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_75_cast_fp16 = transpose(perm = p2c_att_75_perm_0, x = var_4270_cast_fp16_cast_int16)[name = string("transpose_31")];
-            tensor<fp16, [16, 512, 512]> _inversed_4273_cast_fp16 = mul(x = p2c_att_75_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = string("_inversed_4273_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_37_cast_fp16 = add(x = _inversed_4247_cast_fp16, y = _inversed_4273_cast_fp16)[name = string("rel_att_37_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_111_cast_fp16 = add(x = attention_scores_109_cast_fp16, y = rel_att_37_cast_fp16)[name = string("attention_scores_111_cast_fp16")];
-            tensor<int32, [4]> var_4278 = const()[name = string("op_4278"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_113_cast_fp16 = reshape(shape = var_4278, x = attention_scores_111_cast_fp16)[name = string("attention_scores_113_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_441_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_113_cast_fp16, cond = var_339)[name = string("input_441_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_443_cast_fp16 = softmax(axis = var_56, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
-            tensor<int32, [3]> var_4287 = const()[name = string("op_4287"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_4288_cast_fp16 = reshape(shape = var_4287, x = input_443_cast_fp16)[name = string("op_4288_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_75_cast_fp16 = transpose(perm = p2c_att_75_perm_0, x = var_4270_cast_fp16_cast_int16)[name = string("transpose_31")];
+            tensor<fp16, [16, 224, 224]> _inversed_4273_cast_fp16 = mul(x = p2c_att_75_cast_fp16, y = _inversed_4273_y_0_to_fp16)[name = string("_inversed_4273_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_37_cast_fp16 = add(x = _inversed_4247_cast_fp16, y = _inversed_4273_cast_fp16)[name = string("rel_att_37_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_111_cast_fp16 = add(x = attention_scores_109_cast_fp16, y = rel_att_37_cast_fp16)[name = string("attention_scores_111_cast_fp16")];
+            tensor<int32, [4]> var_4278 = const()[name = string("op_4278"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_113_cast_fp16 = reshape(shape = var_4278, x = attention_scores_111_cast_fp16)[name = string("attention_scores_113_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_441_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_113_cast_fp16, cond = var_339)[name = string("input_441_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_443_cast_fp16 = softmax(axis = var_56, x = input_441_cast_fp16)[name = string("input_443_cast_fp16")];
+            tensor<int32, [3]> var_4287 = const()[name = string("op_4287"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_4288_cast_fp16 = reshape(shape = var_4287, x = input_443_cast_fp16)[name = string("op_4288_cast_fp16")];
             bool context_layer_73_transpose_x_0 = const()[name = string("context_layer_73_transpose_x_0"), val = bool(false)];
             bool context_layer_73_transpose_y_0 = const()[name = string("context_layer_73_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_73_cast_fp16 = matmul(transpose_x = context_layer_73_transpose_x_0, transpose_y = context_layer_73_transpose_y_0, x = var_4288_cast_fp16, y = value_layer_37_cast_fp16)[name = string("context_layer_73_cast_fp16")];
-            tensor<int32, [4]> var_4292 = const()[name = string("op_4292"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4293_cast_fp16 = reshape(shape = var_4292, x = context_layer_73_cast_fp16)[name = string("op_4293_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_73_cast_fp16 = matmul(transpose_x = context_layer_73_transpose_x_0, transpose_y = context_layer_73_transpose_y_0, x = var_4288_cast_fp16, y = value_layer_37_cast_fp16)[name = string("context_layer_73_cast_fp16")];
+            tensor<int32, [4]> var_4292 = const()[name = string("op_4292"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4293_cast_fp16 = reshape(shape = var_4292, x = context_layer_73_cast_fp16)[name = string("op_4293_cast_fp16")];
             tensor<int32, [4]> var_4294 = const()[name = string("op_4294"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_4295_cast_fp16 = transpose(perm = var_4294, x = var_4293_cast_fp16)[name = string("transpose_30")];
-            tensor<fp16, [1, 512, 1024]> input_445_cast_fp16 = reshape(shape = var_4299, x = var_4295_cast_fp16)[name = string("input_445_cast_fp16")];
+            tensor<int32, [3]> var_4299 = const()[name = string("op_4299"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_4295_cast_fp16 = transpose(perm = var_4294, x = var_4293_cast_fp16)[name = string("transpose_30")];
+            tensor<fp16, [1, 224, 1024]> input_445_cast_fp16 = reshape(shape = var_4299, x = var_4295_cast_fp16)[name = string("input_445_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_18_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(762848960)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764946176)))];
-            tensor<fp16, [1, 512, 1024]> linear_149_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_attention_output_dense_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_149_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_449_cast_fp16 = add(x = linear_149_cast_fp16, y = input_437_cast_fp16)[name = string("input_449_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_149_cast_fp16 = linear(bias = encoder_encoder_layer_18_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_attention_output_dense_weight_to_fp16, x = input_445_cast_fp16)[name = string("linear_149_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_449_cast_fp16 = add(x = linear_149_cast_fp16, y = input_437_cast_fp16)[name = string("input_449_cast_fp16")];
             tensor<int32, [1]> input_451_axes_0 = const()[name = string("input_451_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764948288)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764950400)))];
-            tensor<fp16, [1, 512, 1024]> input_451_cast_fp16 = layer_norm(axes = input_451_axes_0, beta = encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16, x = input_449_cast_fp16)[name = string("input_451_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_451_cast_fp16 = layer_norm(axes = input_451_axes_0, beta = encoder_encoder_layer_18_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_attention_output_LayerNorm_weight_to_fp16, x = input_449_cast_fp16)[name = string("input_451_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_18_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(764952512)))];
             tensor<fp16, [4096]> encoder_encoder_layer_18_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773341184)))];
-            tensor<fp16, [1, 512, 4096]> linear_150_cast_fp16 = linear(bias = encoder_encoder_layer_18_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_18_intermediate_dense_weight_to_fp16, x = input_451_cast_fp16)[name = string("linear_150_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_150_cast_fp16 = linear(bias = encoder_encoder_layer_18_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_18_intermediate_dense_weight_to_fp16, x = input_451_cast_fp16)[name = string("linear_150_cast_fp16")];
             string input_455_mode_0 = const()[name = string("input_455_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_455_cast_fp16 = gelu(mode = input_455_mode_0, x = linear_150_cast_fp16)[name = string("input_455_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_455_cast_fp16 = gelu(mode = input_455_mode_0, x = linear_150_cast_fp16)[name = string("input_455_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_18_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(773349440)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781738112)))];
-            tensor<fp16, [1, 512, 1024]> linear_151_cast_fp16 = linear(bias = encoder_encoder_layer_18_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_output_dense_weight_to_fp16, x = input_455_cast_fp16)[name = string("linear_151_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_459_cast_fp16 = add(x = linear_151_cast_fp16, y = input_451_cast_fp16)[name = string("input_459_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_151_cast_fp16 = linear(bias = encoder_encoder_layer_18_output_dense_bias_to_fp16, weight = encoder_encoder_layer_18_output_dense_weight_to_fp16, x = input_455_cast_fp16)[name = string("linear_151_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_459_cast_fp16 = add(x = linear_151_cast_fp16, y = input_451_cast_fp16)[name = string("input_459_cast_fp16")];
             tensor<int32, [1]> input_461_axes_0 = const()[name = string("input_461_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781740224)))];
             tensor<fp16, [1024]> encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781742336)))];
-            tensor<fp16, [1, 512, 1024]> input_461_cast_fp16 = layer_norm(axes = input_461_axes_0, beta = encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_461_cast_fp16 = layer_norm(axes = input_461_axes_0, beta = encoder_encoder_layer_18_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_18_output_LayerNorm_weight_to_fp16, x = input_459_cast_fp16)[name = string("input_461_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(781744448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783841664)))];
-            tensor<fp16, [1, 512, 1024]> linear_152_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_152_cast_fp16")];
-            tensor<int32, [4]> var_4350 = const()[name = string("op_4350"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_383_cast_fp16 = reshape(shape = var_4350, x = linear_152_cast_fp16)[name = string("x_383_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_152_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_query_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_152_cast_fp16")];
+            tensor<int32, [4]> var_4350 = const()[name = string("op_4350"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_383_cast_fp16 = reshape(shape = var_4350, x = linear_152_cast_fp16)[name = string("x_383_cast_fp16")];
             tensor<int32, [4]> var_4352 = const()[name = string("op_4352"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4353_cast_fp16 = transpose(perm = var_4352, x = x_383_cast_fp16)[name = string("transpose_29")];
-            tensor<fp16, [16, 512, 64]> query_layer_117_cast_fp16 = reshape(shape = var_4357, x = var_4353_cast_fp16)[name = string("query_layer_117_cast_fp16")];
+            tensor<int32, [3]> var_4357 = const()[name = string("op_4357"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4353_cast_fp16 = transpose(perm = var_4352, x = x_383_cast_fp16)[name = string("transpose_29")];
+            tensor<fp16, [16, 224, 64]> query_layer_117_cast_fp16 = reshape(shape = var_4357, x = var_4353_cast_fp16)[name = string("query_layer_117_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(783843776)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(785940992)))];
-            tensor<fp16, [1, 512, 1024]> linear_153_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_153_cast_fp16")];
-            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_387_cast_fp16 = reshape(shape = var_4364, x = linear_153_cast_fp16)[name = string("x_387_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_153_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_key_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_153_cast_fp16")];
+            tensor<int32, [4]> var_4364 = const()[name = string("op_4364"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_387_cast_fp16 = reshape(shape = var_4364, x = linear_153_cast_fp16)[name = string("x_387_cast_fp16")];
             tensor<int32, [4]> var_4366 = const()[name = string("op_4366"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4371 = const()[name = string("op_4371"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4367_cast_fp16 = transpose(perm = var_4366, x = x_387_cast_fp16)[name = string("transpose_28")];
-            tensor<fp16, [16, 512, 64]> key_layer_39_cast_fp16 = reshape(shape = var_4371, x = var_4367_cast_fp16)[name = string("key_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_4371 = const()[name = string("op_4371"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4367_cast_fp16 = transpose(perm = var_4366, x = x_387_cast_fp16)[name = string("transpose_28")];
+            tensor<fp16, [16, 224, 64]> key_layer_39_cast_fp16 = reshape(shape = var_4371, x = var_4367_cast_fp16)[name = string("key_layer_39_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(785943104)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788040320)))];
-            tensor<fp16, [1, 512, 1024]> linear_154_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_154_cast_fp16")];
-            tensor<int32, [4]> var_4378 = const()[name = string("op_4378"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_391_cast_fp16 = reshape(shape = var_4378, x = linear_154_cast_fp16)[name = string("x_391_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_154_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_19_attention_self_value_proj_weight_to_fp16, x = input_461_cast_fp16)[name = string("linear_154_cast_fp16")];
+            tensor<int32, [4]> var_4378 = const()[name = string("op_4378"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_391_cast_fp16 = reshape(shape = var_4378, x = linear_154_cast_fp16)[name = string("x_391_cast_fp16")];
             tensor<int32, [4]> var_4380 = const()[name = string("op_4380"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4385 = const()[name = string("op_4385"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4381_cast_fp16 = transpose(perm = var_4380, x = x_391_cast_fp16)[name = string("transpose_27")];
-            tensor<fp16, [16, 512, 64]> value_layer_39_cast_fp16 = reshape(shape = var_4385, x = var_4381_cast_fp16)[name = string("value_layer_39_cast_fp16")];
+            tensor<int32, [3]> var_4385 = const()[name = string("op_4385"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4381_cast_fp16 = transpose(perm = var_4380, x = x_391_cast_fp16)[name = string("transpose_27")];
+            tensor<fp16, [16, 224, 64]> value_layer_39_cast_fp16 = reshape(shape = var_4385, x = var_4381_cast_fp16)[name = string("value_layer_39_cast_fp16")];
             tensor<int32, [3]> var_4395_perm_0 = const()[name = string("op_4395_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4397_y_0_to_fp16 = const()[name = string("_inversed_4397_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_4395_cast_fp16 = transpose(perm = var_4395_perm_0, x = key_layer_39_cast_fp16)[name = string("transpose_26")];
-            tensor<fp16, [16, 64, 512]> _inversed_4397_cast_fp16 = mul(x = var_4395_cast_fp16, y = _inversed_4397_y_0_to_fp16)[name = string("_inversed_4397_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_4395_cast_fp16 = transpose(perm = var_4395_perm_0, x = key_layer_39_cast_fp16)[name = string("transpose_26")];
+            tensor<fp16, [16, 64, 224]> _inversed_4397_cast_fp16 = mul(x = var_4395_cast_fp16, y = _inversed_4397_y_0_to_fp16)[name = string("_inversed_4397_cast_fp16")];
             bool attention_scores_115_transpose_x_0 = const()[name = string("attention_scores_115_transpose_x_0"), val = bool(false)];
             bool attention_scores_115_transpose_y_0 = const()[name = string("attention_scores_115_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_115_cast_fp16 = matmul(transpose_x = attention_scores_115_transpose_x_0, transpose_y = attention_scores_115_transpose_y_0, x = query_layer_117_cast_fp16, y = _inversed_4397_cast_fp16)[name = string("attention_scores_115_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_115_cast_fp16 = matmul(transpose_x = attention_scores_115_transpose_x_0, transpose_y = attention_scores_115_transpose_y_0, x = query_layer_117_cast_fp16, y = _inversed_4397_cast_fp16)[name = string("attention_scores_115_cast_fp16")];
             bool c2p_att_77_transpose_x_0 = const()[name = string("c2p_att_77_transpose_x_0"), val = bool(false)];
             bool c2p_att_77_transpose_y_0 = const()[name = string("c2p_att_77_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4453_to_fp16 = const()[name = string("op_4453_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788042432)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_77_cast_fp16 = matmul(transpose_x = c2p_att_77_transpose_x_0, transpose_y = c2p_att_77_transpose_y_0, x = query_layer_117_cast_fp16, y = var_4453_to_fp16)[name = string("c2p_att_77_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_77_cast_fp16 = matmul(transpose_x = c2p_att_77_transpose_x_0, transpose_y = c2p_att_77_transpose_y_0, x = query_layer_117_cast_fp16, y = var_4453_to_fp16)[name = string("c2p_att_77_cast_fp16")];
             bool c2p_att_79_validate_indices_0 = const()[name = string("c2p_att_79_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_79_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_79_validate_indices_0, x = c2p_att_77_cast_fp16)[name = string("c2p_att_79_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_79_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_79_validate_indices_0, x = c2p_att_77_cast_fp16)[name = string("c2p_att_79_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4466_y_0_to_fp16 = const()[name = string("_inversed_4466_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_4466_cast_fp16 = mul(x = c2p_att_79_cast_fp16_cast_int16, y = _inversed_4466_y_0_to_fp16)[name = string("_inversed_4466_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_4466_cast_fp16 = mul(x = c2p_att_79_cast_fp16_cast_int16, y = _inversed_4466_y_0_to_fp16)[name = string("_inversed_4466_cast_fp16")];
             bool p2c_att_77_transpose_x_0 = const()[name = string("p2c_att_77_transpose_x_0"), val = bool(false)];
             bool p2c_att_77_transpose_y_0 = const()[name = string("p2c_att_77_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4481_to_fp16 = const()[name = string("op_4481_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(789091072)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_77_cast_fp16 = matmul(transpose_x = p2c_att_77_transpose_x_0, transpose_y = p2c_att_77_transpose_y_0, x = key_layer_39_cast_fp16, y = var_4481_to_fp16)[name = string("p2c_att_77_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_77_cast_fp16 = matmul(transpose_x = p2c_att_77_transpose_x_0, transpose_y = p2c_att_77_transpose_y_0, x = key_layer_39_cast_fp16, y = var_4481_to_fp16)[name = string("p2c_att_77_cast_fp16")];
             bool var_4489_validate_indices_0 = const()[name = string("op_4489_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_4489_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4489_validate_indices_0, x = p2c_att_77_cast_fp16)[name = string("op_4489_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_4489_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4489_validate_indices_0, x = p2c_att_77_cast_fp16)[name = string("op_4489_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_79_perm_0 = const()[name = string("p2c_att_79_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4492_y_0_to_fp16 = const()[name = string("_inversed_4492_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_79_cast_fp16 = transpose(perm = p2c_att_79_perm_0, x = var_4489_cast_fp16_cast_int16)[name = string("transpose_25")];
-            tensor<fp16, [16, 512, 512]> _inversed_4492_cast_fp16 = mul(x = p2c_att_79_cast_fp16, y = _inversed_4492_y_0_to_fp16)[name = string("_inversed_4492_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_39_cast_fp16 = add(x = _inversed_4466_cast_fp16, y = _inversed_4492_cast_fp16)[name = string("rel_att_39_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_117_cast_fp16 = add(x = attention_scores_115_cast_fp16, y = rel_att_39_cast_fp16)[name = string("attention_scores_117_cast_fp16")];
-            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_119_cast_fp16 = reshape(shape = var_4497, x = attention_scores_117_cast_fp16)[name = string("attention_scores_119_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_465_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_119_cast_fp16, cond = var_339)[name = string("input_465_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_467_cast_fp16 = softmax(axis = var_56, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
-            tensor<int32, [3]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_4507_cast_fp16 = reshape(shape = var_4506, x = input_467_cast_fp16)[name = string("op_4507_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_79_cast_fp16 = transpose(perm = p2c_att_79_perm_0, x = var_4489_cast_fp16_cast_int16)[name = string("transpose_25")];
+            tensor<fp16, [16, 224, 224]> _inversed_4492_cast_fp16 = mul(x = p2c_att_79_cast_fp16, y = _inversed_4492_y_0_to_fp16)[name = string("_inversed_4492_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_39_cast_fp16 = add(x = _inversed_4466_cast_fp16, y = _inversed_4492_cast_fp16)[name = string("rel_att_39_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_117_cast_fp16 = add(x = attention_scores_115_cast_fp16, y = rel_att_39_cast_fp16)[name = string("attention_scores_117_cast_fp16")];
+            tensor<int32, [4]> var_4497 = const()[name = string("op_4497"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_119_cast_fp16 = reshape(shape = var_4497, x = attention_scores_117_cast_fp16)[name = string("attention_scores_119_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_465_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_119_cast_fp16, cond = var_339)[name = string("input_465_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_467_cast_fp16 = softmax(axis = var_56, x = input_465_cast_fp16)[name = string("input_467_cast_fp16")];
+            tensor<int32, [3]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_4507_cast_fp16 = reshape(shape = var_4506, x = input_467_cast_fp16)[name = string("op_4507_cast_fp16")];
             bool context_layer_77_transpose_x_0 = const()[name = string("context_layer_77_transpose_x_0"), val = bool(false)];
             bool context_layer_77_transpose_y_0 = const()[name = string("context_layer_77_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_77_cast_fp16 = matmul(transpose_x = context_layer_77_transpose_x_0, transpose_y = context_layer_77_transpose_y_0, x = var_4507_cast_fp16, y = value_layer_39_cast_fp16)[name = string("context_layer_77_cast_fp16")];
-            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4512_cast_fp16 = reshape(shape = var_4511, x = context_layer_77_cast_fp16)[name = string("op_4512_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_77_cast_fp16 = matmul(transpose_x = context_layer_77_transpose_x_0, transpose_y = context_layer_77_transpose_y_0, x = var_4507_cast_fp16, y = value_layer_39_cast_fp16)[name = string("context_layer_77_cast_fp16")];
+            tensor<int32, [4]> var_4511 = const()[name = string("op_4511"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4512_cast_fp16 = reshape(shape = var_4511, x = context_layer_77_cast_fp16)[name = string("op_4512_cast_fp16")];
             tensor<int32, [4]> var_4513 = const()[name = string("op_4513"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4518 = const()[name = string("op_4518"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_4514_cast_fp16 = transpose(perm = var_4513, x = var_4512_cast_fp16)[name = string("transpose_24")];
-            tensor<fp16, [1, 512, 1024]> input_469_cast_fp16 = reshape(shape = var_4518, x = var_4514_cast_fp16)[name = string("input_469_cast_fp16")];
+            tensor<int32, [3]> var_4518 = const()[name = string("op_4518"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_4514_cast_fp16 = transpose(perm = var_4513, x = var_4512_cast_fp16)[name = string("transpose_24")];
+            tensor<fp16, [1, 224, 1024]> input_469_cast_fp16 = reshape(shape = var_4518, x = var_4514_cast_fp16)[name = string("input_469_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_19_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(790139712)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792236928)))];
-            tensor<fp16, [1, 512, 1024]> linear_157_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_attention_output_dense_weight_to_fp16, x = input_469_cast_fp16)[name = string("linear_157_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_473_cast_fp16 = add(x = linear_157_cast_fp16, y = input_461_cast_fp16)[name = string("input_473_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_157_cast_fp16 = linear(bias = encoder_encoder_layer_19_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_attention_output_dense_weight_to_fp16, x = input_469_cast_fp16)[name = string("linear_157_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_473_cast_fp16 = add(x = linear_157_cast_fp16, y = input_461_cast_fp16)[name = string("input_473_cast_fp16")];
             tensor<int32, [1]> input_475_axes_0 = const()[name = string("input_475_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792239040)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792241152)))];
-            tensor<fp16, [1, 512, 1024]> input_475_cast_fp16 = layer_norm(axes = input_475_axes_0, beta = encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16, x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_475_cast_fp16 = layer_norm(axes = input_475_axes_0, beta = encoder_encoder_layer_19_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_attention_output_LayerNorm_weight_to_fp16, x = input_473_cast_fp16)[name = string("input_475_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_19_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(792243264)))];
             tensor<fp16, [4096]> encoder_encoder_layer_19_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800631936)))];
-            tensor<fp16, [1, 512, 4096]> linear_158_cast_fp16 = linear(bias = encoder_encoder_layer_19_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_19_intermediate_dense_weight_to_fp16, x = input_475_cast_fp16)[name = string("linear_158_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_158_cast_fp16 = linear(bias = encoder_encoder_layer_19_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_19_intermediate_dense_weight_to_fp16, x = input_475_cast_fp16)[name = string("linear_158_cast_fp16")];
             string input_479_mode_0 = const()[name = string("input_479_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_479_cast_fp16 = gelu(mode = input_479_mode_0, x = linear_158_cast_fp16)[name = string("input_479_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_479_cast_fp16 = gelu(mode = input_479_mode_0, x = linear_158_cast_fp16)[name = string("input_479_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_19_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(800640192)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809028864)))];
-            tensor<fp16, [1, 512, 1024]> linear_159_cast_fp16 = linear(bias = encoder_encoder_layer_19_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_output_dense_weight_to_fp16, x = input_479_cast_fp16)[name = string("linear_159_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_483_cast_fp16 = add(x = linear_159_cast_fp16, y = input_475_cast_fp16)[name = string("input_483_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_159_cast_fp16 = linear(bias = encoder_encoder_layer_19_output_dense_bias_to_fp16, weight = encoder_encoder_layer_19_output_dense_weight_to_fp16, x = input_479_cast_fp16)[name = string("linear_159_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_483_cast_fp16 = add(x = linear_159_cast_fp16, y = input_475_cast_fp16)[name = string("input_483_cast_fp16")];
             tensor<int32, [1]> input_485_axes_0 = const()[name = string("input_485_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809030976)))];
             tensor<fp16, [1024]> encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809033088)))];
-            tensor<fp16, [1, 512, 1024]> input_485_cast_fp16 = layer_norm(axes = input_485_axes_0, beta = encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16, x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_485_cast_fp16 = layer_norm(axes = input_485_axes_0, beta = encoder_encoder_layer_19_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_19_output_LayerNorm_weight_to_fp16, x = input_483_cast_fp16)[name = string("input_485_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(809035200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811132416)))];
-            tensor<fp16, [1, 512, 1024]> linear_160_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_160_cast_fp16")];
-            tensor<int32, [4]> var_4569 = const()[name = string("op_4569"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_403_cast_fp16 = reshape(shape = var_4569, x = linear_160_cast_fp16)[name = string("x_403_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_160_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_query_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_160_cast_fp16")];
+            tensor<int32, [4]> var_4569 = const()[name = string("op_4569"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_403_cast_fp16 = reshape(shape = var_4569, x = linear_160_cast_fp16)[name = string("x_403_cast_fp16")];
             tensor<int32, [4]> var_4571 = const()[name = string("op_4571"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4572_cast_fp16 = transpose(perm = var_4571, x = x_403_cast_fp16)[name = string("transpose_23")];
-            tensor<fp16, [16, 512, 64]> query_layer_123_cast_fp16 = reshape(shape = var_4576, x = var_4572_cast_fp16)[name = string("query_layer_123_cast_fp16")];
+            tensor<int32, [3]> var_4576 = const()[name = string("op_4576"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4572_cast_fp16 = transpose(perm = var_4571, x = x_403_cast_fp16)[name = string("transpose_23")];
+            tensor<fp16, [16, 224, 64]> query_layer_123_cast_fp16 = reshape(shape = var_4576, x = var_4572_cast_fp16)[name = string("query_layer_123_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(811134528)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813231744)))];
-            tensor<fp16, [1, 512, 1024]> linear_161_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_161_cast_fp16")];
-            tensor<int32, [4]> var_4583 = const()[name = string("op_4583"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_407_cast_fp16 = reshape(shape = var_4583, x = linear_161_cast_fp16)[name = string("x_407_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_161_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_key_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_161_cast_fp16")];
+            tensor<int32, [4]> var_4583 = const()[name = string("op_4583"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_407_cast_fp16 = reshape(shape = var_4583, x = linear_161_cast_fp16)[name = string("x_407_cast_fp16")];
             tensor<int32, [4]> var_4585 = const()[name = string("op_4585"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4586_cast_fp16 = transpose(perm = var_4585, x = x_407_cast_fp16)[name = string("transpose_22")];
-            tensor<fp16, [16, 512, 64]> key_layer_41_cast_fp16 = reshape(shape = var_4590, x = var_4586_cast_fp16)[name = string("key_layer_41_cast_fp16")];
+            tensor<int32, [3]> var_4590 = const()[name = string("op_4590"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4586_cast_fp16 = transpose(perm = var_4585, x = x_407_cast_fp16)[name = string("transpose_22")];
+            tensor<fp16, [16, 224, 64]> key_layer_41_cast_fp16 = reshape(shape = var_4590, x = var_4586_cast_fp16)[name = string("key_layer_41_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813233856)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815331072)))];
-            tensor<fp16, [1, 512, 1024]> linear_162_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_162_cast_fp16")];
-            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_411_cast_fp16 = reshape(shape = var_4597, x = linear_162_cast_fp16)[name = string("x_411_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_162_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_20_attention_self_value_proj_weight_to_fp16, x = input_485_cast_fp16)[name = string("linear_162_cast_fp16")];
+            tensor<int32, [4]> var_4597 = const()[name = string("op_4597"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_411_cast_fp16 = reshape(shape = var_4597, x = linear_162_cast_fp16)[name = string("x_411_cast_fp16")];
             tensor<int32, [4]> var_4599 = const()[name = string("op_4599"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4600_cast_fp16 = transpose(perm = var_4599, x = x_411_cast_fp16)[name = string("transpose_21")];
-            tensor<fp16, [16, 512, 64]> value_layer_41_cast_fp16 = reshape(shape = var_4604, x = var_4600_cast_fp16)[name = string("value_layer_41_cast_fp16")];
+            tensor<int32, [3]> var_4604 = const()[name = string("op_4604"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4600_cast_fp16 = transpose(perm = var_4599, x = x_411_cast_fp16)[name = string("transpose_21")];
+            tensor<fp16, [16, 224, 64]> value_layer_41_cast_fp16 = reshape(shape = var_4604, x = var_4600_cast_fp16)[name = string("value_layer_41_cast_fp16")];
             tensor<int32, [3]> var_4614_perm_0 = const()[name = string("op_4614_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4616_y_0_to_fp16 = const()[name = string("_inversed_4616_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_4614_cast_fp16 = transpose(perm = var_4614_perm_0, x = key_layer_41_cast_fp16)[name = string("transpose_20")];
-            tensor<fp16, [16, 64, 512]> _inversed_4616_cast_fp16 = mul(x = var_4614_cast_fp16, y = _inversed_4616_y_0_to_fp16)[name = string("_inversed_4616_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_4614_cast_fp16 = transpose(perm = var_4614_perm_0, x = key_layer_41_cast_fp16)[name = string("transpose_20")];
+            tensor<fp16, [16, 64, 224]> _inversed_4616_cast_fp16 = mul(x = var_4614_cast_fp16, y = _inversed_4616_y_0_to_fp16)[name = string("_inversed_4616_cast_fp16")];
             bool attention_scores_121_transpose_x_0 = const()[name = string("attention_scores_121_transpose_x_0"), val = bool(false)];
             bool attention_scores_121_transpose_y_0 = const()[name = string("attention_scores_121_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_121_cast_fp16 = matmul(transpose_x = attention_scores_121_transpose_x_0, transpose_y = attention_scores_121_transpose_y_0, x = query_layer_123_cast_fp16, y = _inversed_4616_cast_fp16)[name = string("attention_scores_121_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_121_cast_fp16 = matmul(transpose_x = attention_scores_121_transpose_x_0, transpose_y = attention_scores_121_transpose_y_0, x = query_layer_123_cast_fp16, y = _inversed_4616_cast_fp16)[name = string("attention_scores_121_cast_fp16")];
             bool c2p_att_81_transpose_x_0 = const()[name = string("c2p_att_81_transpose_x_0"), val = bool(false)];
             bool c2p_att_81_transpose_y_0 = const()[name = string("c2p_att_81_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4672_to_fp16 = const()[name = string("op_4672_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(815333184)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_81_cast_fp16 = matmul(transpose_x = c2p_att_81_transpose_x_0, transpose_y = c2p_att_81_transpose_y_0, x = query_layer_123_cast_fp16, y = var_4672_to_fp16)[name = string("c2p_att_81_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_81_cast_fp16 = matmul(transpose_x = c2p_att_81_transpose_x_0, transpose_y = c2p_att_81_transpose_y_0, x = query_layer_123_cast_fp16, y = var_4672_to_fp16)[name = string("c2p_att_81_cast_fp16")];
             bool c2p_att_83_validate_indices_0 = const()[name = string("c2p_att_83_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_83_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_83_validate_indices_0, x = c2p_att_81_cast_fp16)[name = string("c2p_att_83_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_83_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_83_validate_indices_0, x = c2p_att_81_cast_fp16)[name = string("c2p_att_83_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4685_y_0_to_fp16 = const()[name = string("_inversed_4685_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_4685_cast_fp16 = mul(x = c2p_att_83_cast_fp16_cast_int16, y = _inversed_4685_y_0_to_fp16)[name = string("_inversed_4685_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_4685_cast_fp16 = mul(x = c2p_att_83_cast_fp16_cast_int16, y = _inversed_4685_y_0_to_fp16)[name = string("_inversed_4685_cast_fp16")];
             bool p2c_att_81_transpose_x_0 = const()[name = string("p2c_att_81_transpose_x_0"), val = bool(false)];
             bool p2c_att_81_transpose_y_0 = const()[name = string("p2c_att_81_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4700_to_fp16 = const()[name = string("op_4700_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(816381824)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_81_cast_fp16 = matmul(transpose_x = p2c_att_81_transpose_x_0, transpose_y = p2c_att_81_transpose_y_0, x = key_layer_41_cast_fp16, y = var_4700_to_fp16)[name = string("p2c_att_81_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_81_cast_fp16 = matmul(transpose_x = p2c_att_81_transpose_x_0, transpose_y = p2c_att_81_transpose_y_0, x = key_layer_41_cast_fp16, y = var_4700_to_fp16)[name = string("p2c_att_81_cast_fp16")];
             bool var_4708_validate_indices_0 = const()[name = string("op_4708_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_4708_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4708_validate_indices_0, x = p2c_att_81_cast_fp16)[name = string("op_4708_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_4708_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4708_validate_indices_0, x = p2c_att_81_cast_fp16)[name = string("op_4708_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_83_perm_0 = const()[name = string("p2c_att_83_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4711_y_0_to_fp16 = const()[name = string("_inversed_4711_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_83_cast_fp16 = transpose(perm = p2c_att_83_perm_0, x = var_4708_cast_fp16_cast_int16)[name = string("transpose_19")];
-            tensor<fp16, [16, 512, 512]> _inversed_4711_cast_fp16 = mul(x = p2c_att_83_cast_fp16, y = _inversed_4711_y_0_to_fp16)[name = string("_inversed_4711_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_41_cast_fp16 = add(x = _inversed_4685_cast_fp16, y = _inversed_4711_cast_fp16)[name = string("rel_att_41_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_123_cast_fp16 = add(x = attention_scores_121_cast_fp16, y = rel_att_41_cast_fp16)[name = string("attention_scores_123_cast_fp16")];
-            tensor<int32, [4]> var_4716 = const()[name = string("op_4716"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_125_cast_fp16 = reshape(shape = var_4716, x = attention_scores_123_cast_fp16)[name = string("attention_scores_125_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_489_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_125_cast_fp16, cond = var_339)[name = string("input_489_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_491_cast_fp16 = softmax(axis = var_56, x = input_489_cast_fp16)[name = string("input_491_cast_fp16")];
-            tensor<int32, [3]> var_4725 = const()[name = string("op_4725"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_4726_cast_fp16 = reshape(shape = var_4725, x = input_491_cast_fp16)[name = string("op_4726_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_83_cast_fp16 = transpose(perm = p2c_att_83_perm_0, x = var_4708_cast_fp16_cast_int16)[name = string("transpose_19")];
+            tensor<fp16, [16, 224, 224]> _inversed_4711_cast_fp16 = mul(x = p2c_att_83_cast_fp16, y = _inversed_4711_y_0_to_fp16)[name = string("_inversed_4711_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_41_cast_fp16 = add(x = _inversed_4685_cast_fp16, y = _inversed_4711_cast_fp16)[name = string("rel_att_41_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_123_cast_fp16 = add(x = attention_scores_121_cast_fp16, y = rel_att_41_cast_fp16)[name = string("attention_scores_123_cast_fp16")];
+            tensor<int32, [4]> var_4716 = const()[name = string("op_4716"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_125_cast_fp16 = reshape(shape = var_4716, x = attention_scores_123_cast_fp16)[name = string("attention_scores_125_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_489_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_125_cast_fp16, cond = var_339)[name = string("input_489_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_491_cast_fp16 = softmax(axis = var_56, x = input_489_cast_fp16)[name = string("input_491_cast_fp16")];
+            tensor<int32, [3]> var_4725 = const()[name = string("op_4725"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_4726_cast_fp16 = reshape(shape = var_4725, x = input_491_cast_fp16)[name = string("op_4726_cast_fp16")];
             bool context_layer_81_transpose_x_0 = const()[name = string("context_layer_81_transpose_x_0"), val = bool(false)];
             bool context_layer_81_transpose_y_0 = const()[name = string("context_layer_81_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_81_cast_fp16 = matmul(transpose_x = context_layer_81_transpose_x_0, transpose_y = context_layer_81_transpose_y_0, x = var_4726_cast_fp16, y = value_layer_41_cast_fp16)[name = string("context_layer_81_cast_fp16")];
-            tensor<int32, [4]> var_4730 = const()[name = string("op_4730"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4731_cast_fp16 = reshape(shape = var_4730, x = context_layer_81_cast_fp16)[name = string("op_4731_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_81_cast_fp16 = matmul(transpose_x = context_layer_81_transpose_x_0, transpose_y = context_layer_81_transpose_y_0, x = var_4726_cast_fp16, y = value_layer_41_cast_fp16)[name = string("context_layer_81_cast_fp16")];
+            tensor<int32, [4]> var_4730 = const()[name = string("op_4730"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4731_cast_fp16 = reshape(shape = var_4730, x = context_layer_81_cast_fp16)[name = string("op_4731_cast_fp16")];
             tensor<int32, [4]> var_4732 = const()[name = string("op_4732"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4737 = const()[name = string("op_4737"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_4733_cast_fp16 = transpose(perm = var_4732, x = var_4731_cast_fp16)[name = string("transpose_18")];
-            tensor<fp16, [1, 512, 1024]> input_493_cast_fp16 = reshape(shape = var_4737, x = var_4733_cast_fp16)[name = string("input_493_cast_fp16")];
+            tensor<int32, [3]> var_4737 = const()[name = string("op_4737"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_4733_cast_fp16 = transpose(perm = var_4732, x = var_4731_cast_fp16)[name = string("transpose_18")];
+            tensor<fp16, [1, 224, 1024]> input_493_cast_fp16 = reshape(shape = var_4737, x = var_4733_cast_fp16)[name = string("input_493_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_20_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(817430464)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819527680)))];
-            tensor<fp16, [1, 512, 1024]> linear_165_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_attention_output_dense_weight_to_fp16, x = input_493_cast_fp16)[name = string("linear_165_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_497_cast_fp16 = add(x = linear_165_cast_fp16, y = input_485_cast_fp16)[name = string("input_497_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_165_cast_fp16 = linear(bias = encoder_encoder_layer_20_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_attention_output_dense_weight_to_fp16, x = input_493_cast_fp16)[name = string("linear_165_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_497_cast_fp16 = add(x = linear_165_cast_fp16, y = input_485_cast_fp16)[name = string("input_497_cast_fp16")];
             tensor<int32, [1]> input_499_axes_0 = const()[name = string("input_499_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819529792)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819531904)))];
-            tensor<fp16, [1, 512, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, beta = encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_499_cast_fp16 = layer_norm(axes = input_499_axes_0, beta = encoder_encoder_layer_20_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_attention_output_LayerNorm_weight_to_fp16, x = input_497_cast_fp16)[name = string("input_499_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_20_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(819534016)))];
             tensor<fp16, [4096]> encoder_encoder_layer_20_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827922688)))];
-            tensor<fp16, [1, 512, 4096]> linear_166_cast_fp16 = linear(bias = encoder_encoder_layer_20_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_20_intermediate_dense_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_166_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_166_cast_fp16 = linear(bias = encoder_encoder_layer_20_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_20_intermediate_dense_weight_to_fp16, x = input_499_cast_fp16)[name = string("linear_166_cast_fp16")];
             string input_503_mode_0 = const()[name = string("input_503_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_503_cast_fp16 = gelu(mode = input_503_mode_0, x = linear_166_cast_fp16)[name = string("input_503_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_503_cast_fp16 = gelu(mode = input_503_mode_0, x = linear_166_cast_fp16)[name = string("input_503_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_20_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(827930944)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836319616)))];
-            tensor<fp16, [1, 512, 1024]> linear_167_cast_fp16 = linear(bias = encoder_encoder_layer_20_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_output_dense_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_167_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_507_cast_fp16 = add(x = linear_167_cast_fp16, y = input_499_cast_fp16)[name = string("input_507_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_167_cast_fp16 = linear(bias = encoder_encoder_layer_20_output_dense_bias_to_fp16, weight = encoder_encoder_layer_20_output_dense_weight_to_fp16, x = input_503_cast_fp16)[name = string("linear_167_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_507_cast_fp16 = add(x = linear_167_cast_fp16, y = input_499_cast_fp16)[name = string("input_507_cast_fp16")];
             tensor<int32, [1]> input_509_axes_0 = const()[name = string("input_509_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836321728)))];
             tensor<fp16, [1024]> encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836323840)))];
-            tensor<fp16, [1, 512, 1024]> input_509_cast_fp16 = layer_norm(axes = input_509_axes_0, beta = encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_509_cast_fp16 = layer_norm(axes = input_509_axes_0, beta = encoder_encoder_layer_20_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_20_output_LayerNorm_weight_to_fp16, x = input_507_cast_fp16)[name = string("input_509_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836325952)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(838423168)))];
-            tensor<fp16, [1, 512, 1024]> linear_168_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_168_cast_fp16")];
-            tensor<int32, [4]> var_4788 = const()[name = string("op_4788"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_423_cast_fp16 = reshape(shape = var_4788, x = linear_168_cast_fp16)[name = string("x_423_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_168_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_query_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_168_cast_fp16")];
+            tensor<int32, [4]> var_4788 = const()[name = string("op_4788"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_423_cast_fp16 = reshape(shape = var_4788, x = linear_168_cast_fp16)[name = string("x_423_cast_fp16")];
             tensor<int32, [4]> var_4790 = const()[name = string("op_4790"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4795 = const()[name = string("op_4795"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4791_cast_fp16 = transpose(perm = var_4790, x = x_423_cast_fp16)[name = string("transpose_17")];
-            tensor<fp16, [16, 512, 64]> query_layer_129_cast_fp16 = reshape(shape = var_4795, x = var_4791_cast_fp16)[name = string("query_layer_129_cast_fp16")];
+            tensor<int32, [3]> var_4795 = const()[name = string("op_4795"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4791_cast_fp16 = transpose(perm = var_4790, x = x_423_cast_fp16)[name = string("transpose_17")];
+            tensor<fp16, [16, 224, 64]> query_layer_129_cast_fp16 = reshape(shape = var_4795, x = var_4791_cast_fp16)[name = string("query_layer_129_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(838425280)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(840522496)))];
-            tensor<fp16, [1, 512, 1024]> linear_169_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_169_cast_fp16")];
-            tensor<int32, [4]> var_4802 = const()[name = string("op_4802"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_427_cast_fp16 = reshape(shape = var_4802, x = linear_169_cast_fp16)[name = string("x_427_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_169_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_key_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_169_cast_fp16")];
+            tensor<int32, [4]> var_4802 = const()[name = string("op_4802"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_427_cast_fp16 = reshape(shape = var_4802, x = linear_169_cast_fp16)[name = string("x_427_cast_fp16")];
             tensor<int32, [4]> var_4804 = const()[name = string("op_4804"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4805_cast_fp16 = transpose(perm = var_4804, x = x_427_cast_fp16)[name = string("transpose_16")];
-            tensor<fp16, [16, 512, 64]> key_layer_43_cast_fp16 = reshape(shape = var_4809, x = var_4805_cast_fp16)[name = string("key_layer_43_cast_fp16")];
+            tensor<int32, [3]> var_4809 = const()[name = string("op_4809"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4805_cast_fp16 = transpose(perm = var_4804, x = x_427_cast_fp16)[name = string("transpose_16")];
+            tensor<fp16, [16, 224, 64]> key_layer_43_cast_fp16 = reshape(shape = var_4809, x = var_4805_cast_fp16)[name = string("key_layer_43_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(840524608)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842621824)))];
-            tensor<fp16, [1, 512, 1024]> linear_170_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_170_cast_fp16")];
-            tensor<int32, [4]> var_4816 = const()[name = string("op_4816"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_431_cast_fp16 = reshape(shape = var_4816, x = linear_170_cast_fp16)[name = string("x_431_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_170_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_21_attention_self_value_proj_weight_to_fp16, x = input_509_cast_fp16)[name = string("linear_170_cast_fp16")];
+            tensor<int32, [4]> var_4816 = const()[name = string("op_4816"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_431_cast_fp16 = reshape(shape = var_4816, x = linear_170_cast_fp16)[name = string("x_431_cast_fp16")];
             tensor<int32, [4]> var_4818 = const()[name = string("op_4818"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4823 = const()[name = string("op_4823"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4819_cast_fp16 = transpose(perm = var_4818, x = x_431_cast_fp16)[name = string("transpose_15")];
-            tensor<fp16, [16, 512, 64]> value_layer_43_cast_fp16 = reshape(shape = var_4823, x = var_4819_cast_fp16)[name = string("value_layer_43_cast_fp16")];
+            tensor<int32, [3]> var_4823 = const()[name = string("op_4823"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4819_cast_fp16 = transpose(perm = var_4818, x = x_431_cast_fp16)[name = string("transpose_15")];
+            tensor<fp16, [16, 224, 64]> value_layer_43_cast_fp16 = reshape(shape = var_4823, x = var_4819_cast_fp16)[name = string("value_layer_43_cast_fp16")];
             tensor<int32, [3]> var_4833_perm_0 = const()[name = string("op_4833_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4835_y_0_to_fp16 = const()[name = string("_inversed_4835_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_4833_cast_fp16 = transpose(perm = var_4833_perm_0, x = key_layer_43_cast_fp16)[name = string("transpose_14")];
-            tensor<fp16, [16, 64, 512]> _inversed_4835_cast_fp16 = mul(x = var_4833_cast_fp16, y = _inversed_4835_y_0_to_fp16)[name = string("_inversed_4835_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_4833_cast_fp16 = transpose(perm = var_4833_perm_0, x = key_layer_43_cast_fp16)[name = string("transpose_14")];
+            tensor<fp16, [16, 64, 224]> _inversed_4835_cast_fp16 = mul(x = var_4833_cast_fp16, y = _inversed_4835_y_0_to_fp16)[name = string("_inversed_4835_cast_fp16")];
             bool attention_scores_127_transpose_x_0 = const()[name = string("attention_scores_127_transpose_x_0"), val = bool(false)];
             bool attention_scores_127_transpose_y_0 = const()[name = string("attention_scores_127_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_127_cast_fp16 = matmul(transpose_x = attention_scores_127_transpose_x_0, transpose_y = attention_scores_127_transpose_y_0, x = query_layer_129_cast_fp16, y = _inversed_4835_cast_fp16)[name = string("attention_scores_127_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_127_cast_fp16 = matmul(transpose_x = attention_scores_127_transpose_x_0, transpose_y = attention_scores_127_transpose_y_0, x = query_layer_129_cast_fp16, y = _inversed_4835_cast_fp16)[name = string("attention_scores_127_cast_fp16")];
             bool c2p_att_85_transpose_x_0 = const()[name = string("c2p_att_85_transpose_x_0"), val = bool(false)];
             bool c2p_att_85_transpose_y_0 = const()[name = string("c2p_att_85_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4891_to_fp16 = const()[name = string("op_4891_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(842623936)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_85_cast_fp16 = matmul(transpose_x = c2p_att_85_transpose_x_0, transpose_y = c2p_att_85_transpose_y_0, x = query_layer_129_cast_fp16, y = var_4891_to_fp16)[name = string("c2p_att_85_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_85_cast_fp16 = matmul(transpose_x = c2p_att_85_transpose_x_0, transpose_y = c2p_att_85_transpose_y_0, x = query_layer_129_cast_fp16, y = var_4891_to_fp16)[name = string("c2p_att_85_cast_fp16")];
             bool c2p_att_87_validate_indices_0 = const()[name = string("c2p_att_87_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_87_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_87_validate_indices_0, x = c2p_att_85_cast_fp16)[name = string("c2p_att_87_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_87_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_87_validate_indices_0, x = c2p_att_85_cast_fp16)[name = string("c2p_att_87_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_4904_y_0_to_fp16 = const()[name = string("_inversed_4904_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_4904_cast_fp16 = mul(x = c2p_att_87_cast_fp16_cast_int16, y = _inversed_4904_y_0_to_fp16)[name = string("_inversed_4904_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_4904_cast_fp16 = mul(x = c2p_att_87_cast_fp16_cast_int16, y = _inversed_4904_y_0_to_fp16)[name = string("_inversed_4904_cast_fp16")];
             bool p2c_att_85_transpose_x_0 = const()[name = string("p2c_att_85_transpose_x_0"), val = bool(false)];
             bool p2c_att_85_transpose_y_0 = const()[name = string("p2c_att_85_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_4919_to_fp16 = const()[name = string("op_4919_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(843672576)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_85_cast_fp16 = matmul(transpose_x = p2c_att_85_transpose_x_0, transpose_y = p2c_att_85_transpose_y_0, x = key_layer_43_cast_fp16, y = var_4919_to_fp16)[name = string("p2c_att_85_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_85_cast_fp16 = matmul(transpose_x = p2c_att_85_transpose_x_0, transpose_y = p2c_att_85_transpose_y_0, x = key_layer_43_cast_fp16, y = var_4919_to_fp16)[name = string("p2c_att_85_cast_fp16")];
             bool var_4927_validate_indices_0 = const()[name = string("op_4927_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_4927_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4927_validate_indices_0, x = p2c_att_85_cast_fp16)[name = string("op_4927_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_4927_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_4927_validate_indices_0, x = p2c_att_85_cast_fp16)[name = string("op_4927_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_87_perm_0 = const()[name = string("p2c_att_87_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_4930_y_0_to_fp16 = const()[name = string("_inversed_4930_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_87_cast_fp16 = transpose(perm = p2c_att_87_perm_0, x = var_4927_cast_fp16_cast_int16)[name = string("transpose_13")];
-            tensor<fp16, [16, 512, 512]> _inversed_4930_cast_fp16 = mul(x = p2c_att_87_cast_fp16, y = _inversed_4930_y_0_to_fp16)[name = string("_inversed_4930_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_43_cast_fp16 = add(x = _inversed_4904_cast_fp16, y = _inversed_4930_cast_fp16)[name = string("rel_att_43_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_129_cast_fp16 = add(x = attention_scores_127_cast_fp16, y = rel_att_43_cast_fp16)[name = string("attention_scores_129_cast_fp16")];
-            tensor<int32, [4]> var_4935 = const()[name = string("op_4935"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_131_cast_fp16 = reshape(shape = var_4935, x = attention_scores_129_cast_fp16)[name = string("attention_scores_131_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_513_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_131_cast_fp16, cond = var_339)[name = string("input_513_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_515_cast_fp16 = softmax(axis = var_56, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
-            tensor<int32, [3]> var_4944 = const()[name = string("op_4944"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_4945_cast_fp16 = reshape(shape = var_4944, x = input_515_cast_fp16)[name = string("op_4945_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_87_cast_fp16 = transpose(perm = p2c_att_87_perm_0, x = var_4927_cast_fp16_cast_int16)[name = string("transpose_13")];
+            tensor<fp16, [16, 224, 224]> _inversed_4930_cast_fp16 = mul(x = p2c_att_87_cast_fp16, y = _inversed_4930_y_0_to_fp16)[name = string("_inversed_4930_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_43_cast_fp16 = add(x = _inversed_4904_cast_fp16, y = _inversed_4930_cast_fp16)[name = string("rel_att_43_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_129_cast_fp16 = add(x = attention_scores_127_cast_fp16, y = rel_att_43_cast_fp16)[name = string("attention_scores_129_cast_fp16")];
+            tensor<int32, [4]> var_4935 = const()[name = string("op_4935"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_131_cast_fp16 = reshape(shape = var_4935, x = attention_scores_129_cast_fp16)[name = string("attention_scores_131_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_513_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_131_cast_fp16, cond = var_339)[name = string("input_513_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_515_cast_fp16 = softmax(axis = var_56, x = input_513_cast_fp16)[name = string("input_515_cast_fp16")];
+            tensor<int32, [3]> var_4944 = const()[name = string("op_4944"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_4945_cast_fp16 = reshape(shape = var_4944, x = input_515_cast_fp16)[name = string("op_4945_cast_fp16")];
             bool context_layer_85_transpose_x_0 = const()[name = string("context_layer_85_transpose_x_0"), val = bool(false)];
             bool context_layer_85_transpose_y_0 = const()[name = string("context_layer_85_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_85_cast_fp16 = matmul(transpose_x = context_layer_85_transpose_x_0, transpose_y = context_layer_85_transpose_y_0, x = var_4945_cast_fp16, y = value_layer_43_cast_fp16)[name = string("context_layer_85_cast_fp16")];
-            tensor<int32, [4]> var_4949 = const()[name = string("op_4949"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_4950_cast_fp16 = reshape(shape = var_4949, x = context_layer_85_cast_fp16)[name = string("op_4950_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_85_cast_fp16 = matmul(transpose_x = context_layer_85_transpose_x_0, transpose_y = context_layer_85_transpose_y_0, x = var_4945_cast_fp16, y = value_layer_43_cast_fp16)[name = string("context_layer_85_cast_fp16")];
+            tensor<int32, [4]> var_4949 = const()[name = string("op_4949"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_4950_cast_fp16 = reshape(shape = var_4949, x = context_layer_85_cast_fp16)[name = string("op_4950_cast_fp16")];
             tensor<int32, [4]> var_4951 = const()[name = string("op_4951"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_4956 = const()[name = string("op_4956"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_4952_cast_fp16 = transpose(perm = var_4951, x = var_4950_cast_fp16)[name = string("transpose_12")];
-            tensor<fp16, [1, 512, 1024]> input_517_cast_fp16 = reshape(shape = var_4956, x = var_4952_cast_fp16)[name = string("input_517_cast_fp16")];
+            tensor<int32, [3]> var_4956 = const()[name = string("op_4956"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_4952_cast_fp16 = transpose(perm = var_4951, x = var_4950_cast_fp16)[name = string("transpose_12")];
+            tensor<fp16, [1, 224, 1024]> input_517_cast_fp16 = reshape(shape = var_4956, x = var_4952_cast_fp16)[name = string("input_517_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_21_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(844721216)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846818432)))];
-            tensor<fp16, [1, 512, 1024]> linear_173_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_attention_output_dense_weight_to_fp16, x = input_517_cast_fp16)[name = string("linear_173_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_521_cast_fp16 = add(x = linear_173_cast_fp16, y = input_509_cast_fp16)[name = string("input_521_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_173_cast_fp16 = linear(bias = encoder_encoder_layer_21_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_attention_output_dense_weight_to_fp16, x = input_517_cast_fp16)[name = string("linear_173_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_521_cast_fp16 = add(x = linear_173_cast_fp16, y = input_509_cast_fp16)[name = string("input_521_cast_fp16")];
             tensor<int32, [1]> input_523_axes_0 = const()[name = string("input_523_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846820544)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846822656)))];
-            tensor<fp16, [1, 512, 1024]> input_523_cast_fp16 = layer_norm(axes = input_523_axes_0, beta = encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_523_cast_fp16 = layer_norm(axes = input_523_axes_0, beta = encoder_encoder_layer_21_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_attention_output_LayerNorm_weight_to_fp16, x = input_521_cast_fp16)[name = string("input_523_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_21_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(846824768)))];
             tensor<fp16, [4096]> encoder_encoder_layer_21_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855213440)))];
-            tensor<fp16, [1, 512, 4096]> linear_174_cast_fp16 = linear(bias = encoder_encoder_layer_21_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_21_intermediate_dense_weight_to_fp16, x = input_523_cast_fp16)[name = string("linear_174_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_174_cast_fp16 = linear(bias = encoder_encoder_layer_21_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_21_intermediate_dense_weight_to_fp16, x = input_523_cast_fp16)[name = string("linear_174_cast_fp16")];
             string input_527_mode_0 = const()[name = string("input_527_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_527_cast_fp16 = gelu(mode = input_527_mode_0, x = linear_174_cast_fp16)[name = string("input_527_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_527_cast_fp16 = gelu(mode = input_527_mode_0, x = linear_174_cast_fp16)[name = string("input_527_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_21_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855221696)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863610368)))];
-            tensor<fp16, [1, 512, 1024]> linear_175_cast_fp16 = linear(bias = encoder_encoder_layer_21_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_output_dense_weight_to_fp16, x = input_527_cast_fp16)[name = string("linear_175_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_531_cast_fp16 = add(x = linear_175_cast_fp16, y = input_523_cast_fp16)[name = string("input_531_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_175_cast_fp16 = linear(bias = encoder_encoder_layer_21_output_dense_bias_to_fp16, weight = encoder_encoder_layer_21_output_dense_weight_to_fp16, x = input_527_cast_fp16)[name = string("linear_175_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_531_cast_fp16 = add(x = linear_175_cast_fp16, y = input_523_cast_fp16)[name = string("input_531_cast_fp16")];
             tensor<int32, [1]> input_533_axes_0 = const()[name = string("input_533_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863612480)))];
             tensor<fp16, [1024]> encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863614592)))];
-            tensor<fp16, [1, 512, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_533_cast_fp16 = layer_norm(axes = input_533_axes_0, beta = encoder_encoder_layer_21_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_21_output_LayerNorm_weight_to_fp16, x = input_531_cast_fp16)[name = string("input_533_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(863616704)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(865713920)))];
-            tensor<fp16, [1, 512, 1024]> linear_176_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_176_cast_fp16")];
-            tensor<int32, [4]> var_5007 = const()[name = string("op_5007"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_443_cast_fp16 = reshape(shape = var_5007, x = linear_176_cast_fp16)[name = string("x_443_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_176_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_query_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_176_cast_fp16")];
+            tensor<int32, [4]> var_5007 = const()[name = string("op_5007"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_443_cast_fp16 = reshape(shape = var_5007, x = linear_176_cast_fp16)[name = string("x_443_cast_fp16")];
             tensor<int32, [4]> var_5009 = const()[name = string("op_5009"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5014 = const()[name = string("op_5014"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5010_cast_fp16 = transpose(perm = var_5009, x = x_443_cast_fp16)[name = string("transpose_11")];
-            tensor<fp16, [16, 512, 64]> query_layer_135_cast_fp16 = reshape(shape = var_5014, x = var_5010_cast_fp16)[name = string("query_layer_135_cast_fp16")];
+            tensor<int32, [3]> var_5014 = const()[name = string("op_5014"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5010_cast_fp16 = transpose(perm = var_5009, x = x_443_cast_fp16)[name = string("transpose_11")];
+            tensor<fp16, [16, 224, 64]> query_layer_135_cast_fp16 = reshape(shape = var_5014, x = var_5010_cast_fp16)[name = string("query_layer_135_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(865716032)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867813248)))];
-            tensor<fp16, [1, 512, 1024]> linear_177_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_177_cast_fp16")];
-            tensor<int32, [4]> var_5021 = const()[name = string("op_5021"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_447_cast_fp16 = reshape(shape = var_5021, x = linear_177_cast_fp16)[name = string("x_447_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_177_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_key_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_177_cast_fp16")];
+            tensor<int32, [4]> var_5021 = const()[name = string("op_5021"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_447_cast_fp16 = reshape(shape = var_5021, x = linear_177_cast_fp16)[name = string("x_447_cast_fp16")];
             tensor<int32, [4]> var_5023 = const()[name = string("op_5023"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5028 = const()[name = string("op_5028"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5024_cast_fp16 = transpose(perm = var_5023, x = x_447_cast_fp16)[name = string("transpose_10")];
-            tensor<fp16, [16, 512, 64]> key_layer_45_cast_fp16 = reshape(shape = var_5028, x = var_5024_cast_fp16)[name = string("key_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_5028 = const()[name = string("op_5028"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5024_cast_fp16 = transpose(perm = var_5023, x = x_447_cast_fp16)[name = string("transpose_10")];
+            tensor<fp16, [16, 224, 64]> key_layer_45_cast_fp16 = reshape(shape = var_5028, x = var_5024_cast_fp16)[name = string("key_layer_45_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(867815360)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(869912576)))];
-            tensor<fp16, [1, 512, 1024]> linear_178_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_178_cast_fp16")];
-            tensor<int32, [4]> var_5035 = const()[name = string("op_5035"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_451_cast_fp16 = reshape(shape = var_5035, x = linear_178_cast_fp16)[name = string("x_451_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_178_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_22_attention_self_value_proj_weight_to_fp16, x = input_533_cast_fp16)[name = string("linear_178_cast_fp16")];
+            tensor<int32, [4]> var_5035 = const()[name = string("op_5035"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_451_cast_fp16 = reshape(shape = var_5035, x = linear_178_cast_fp16)[name = string("x_451_cast_fp16")];
             tensor<int32, [4]> var_5037 = const()[name = string("op_5037"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5042 = const()[name = string("op_5042"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5038_cast_fp16 = transpose(perm = var_5037, x = x_451_cast_fp16)[name = string("transpose_9")];
-            tensor<fp16, [16, 512, 64]> value_layer_45_cast_fp16 = reshape(shape = var_5042, x = var_5038_cast_fp16)[name = string("value_layer_45_cast_fp16")];
+            tensor<int32, [3]> var_5042 = const()[name = string("op_5042"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5038_cast_fp16 = transpose(perm = var_5037, x = x_451_cast_fp16)[name = string("transpose_9")];
+            tensor<fp16, [16, 224, 64]> value_layer_45_cast_fp16 = reshape(shape = var_5042, x = var_5038_cast_fp16)[name = string("value_layer_45_cast_fp16")];
             tensor<int32, [3]> var_5052_perm_0 = const()[name = string("op_5052_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5054_y_0_to_fp16 = const()[name = string("_inversed_5054_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_5052_cast_fp16 = transpose(perm = var_5052_perm_0, x = key_layer_45_cast_fp16)[name = string("transpose_8")];
-            tensor<fp16, [16, 64, 512]> _inversed_5054_cast_fp16 = mul(x = var_5052_cast_fp16, y = _inversed_5054_y_0_to_fp16)[name = string("_inversed_5054_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_5052_cast_fp16 = transpose(perm = var_5052_perm_0, x = key_layer_45_cast_fp16)[name = string("transpose_8")];
+            tensor<fp16, [16, 64, 224]> _inversed_5054_cast_fp16 = mul(x = var_5052_cast_fp16, y = _inversed_5054_y_0_to_fp16)[name = string("_inversed_5054_cast_fp16")];
             bool attention_scores_133_transpose_x_0 = const()[name = string("attention_scores_133_transpose_x_0"), val = bool(false)];
             bool attention_scores_133_transpose_y_0 = const()[name = string("attention_scores_133_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_133_cast_fp16 = matmul(transpose_x = attention_scores_133_transpose_x_0, transpose_y = attention_scores_133_transpose_y_0, x = query_layer_135_cast_fp16, y = _inversed_5054_cast_fp16)[name = string("attention_scores_133_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_133_cast_fp16 = matmul(transpose_x = attention_scores_133_transpose_x_0, transpose_y = attention_scores_133_transpose_y_0, x = query_layer_135_cast_fp16, y = _inversed_5054_cast_fp16)[name = string("attention_scores_133_cast_fp16")];
             bool c2p_att_89_transpose_x_0 = const()[name = string("c2p_att_89_transpose_x_0"), val = bool(false)];
             bool c2p_att_89_transpose_y_0 = const()[name = string("c2p_att_89_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5110_to_fp16 = const()[name = string("op_5110_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(869914688)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_89_cast_fp16 = matmul(transpose_x = c2p_att_89_transpose_x_0, transpose_y = c2p_att_89_transpose_y_0, x = query_layer_135_cast_fp16, y = var_5110_to_fp16)[name = string("c2p_att_89_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_89_cast_fp16 = matmul(transpose_x = c2p_att_89_transpose_x_0, transpose_y = c2p_att_89_transpose_y_0, x = query_layer_135_cast_fp16, y = var_5110_to_fp16)[name = string("c2p_att_89_cast_fp16")];
             bool c2p_att_91_validate_indices_0 = const()[name = string("c2p_att_91_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_91_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_91_validate_indices_0, x = c2p_att_89_cast_fp16)[name = string("c2p_att_91_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_91_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_91_validate_indices_0, x = c2p_att_89_cast_fp16)[name = string("c2p_att_91_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_5123_y_0_to_fp16 = const()[name = string("_inversed_5123_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_5123_cast_fp16 = mul(x = c2p_att_91_cast_fp16_cast_int16, y = _inversed_5123_y_0_to_fp16)[name = string("_inversed_5123_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_5123_cast_fp16 = mul(x = c2p_att_91_cast_fp16_cast_int16, y = _inversed_5123_y_0_to_fp16)[name = string("_inversed_5123_cast_fp16")];
             bool p2c_att_89_transpose_x_0 = const()[name = string("p2c_att_89_transpose_x_0"), val = bool(false)];
             bool p2c_att_89_transpose_y_0 = const()[name = string("p2c_att_89_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5138_to_fp16 = const()[name = string("op_5138_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(870963328)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_89_cast_fp16 = matmul(transpose_x = p2c_att_89_transpose_x_0, transpose_y = p2c_att_89_transpose_y_0, x = key_layer_45_cast_fp16, y = var_5138_to_fp16)[name = string("p2c_att_89_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_89_cast_fp16 = matmul(transpose_x = p2c_att_89_transpose_x_0, transpose_y = p2c_att_89_transpose_y_0, x = key_layer_45_cast_fp16, y = var_5138_to_fp16)[name = string("p2c_att_89_cast_fp16")];
             bool var_5146_validate_indices_0 = const()[name = string("op_5146_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_5146_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5146_validate_indices_0, x = p2c_att_89_cast_fp16)[name = string("op_5146_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_5146_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5146_validate_indices_0, x = p2c_att_89_cast_fp16)[name = string("op_5146_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_91_perm_0 = const()[name = string("p2c_att_91_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5149_y_0_to_fp16 = const()[name = string("_inversed_5149_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_91_cast_fp16 = transpose(perm = p2c_att_91_perm_0, x = var_5146_cast_fp16_cast_int16)[name = string("transpose_7")];
-            tensor<fp16, [16, 512, 512]> _inversed_5149_cast_fp16 = mul(x = p2c_att_91_cast_fp16, y = _inversed_5149_y_0_to_fp16)[name = string("_inversed_5149_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_45_cast_fp16 = add(x = _inversed_5123_cast_fp16, y = _inversed_5149_cast_fp16)[name = string("rel_att_45_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_135_cast_fp16 = add(x = attention_scores_133_cast_fp16, y = rel_att_45_cast_fp16)[name = string("attention_scores_135_cast_fp16")];
-            tensor<int32, [4]> var_5154 = const()[name = string("op_5154"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_137_cast_fp16 = reshape(shape = var_5154, x = attention_scores_135_cast_fp16)[name = string("attention_scores_137_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_537_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_137_cast_fp16, cond = var_339)[name = string("input_537_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_539_cast_fp16 = softmax(axis = var_56, x = input_537_cast_fp16)[name = string("input_539_cast_fp16")];
-            tensor<int32, [3]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_5164_cast_fp16 = reshape(shape = var_5163, x = input_539_cast_fp16)[name = string("op_5164_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_91_cast_fp16 = transpose(perm = p2c_att_91_perm_0, x = var_5146_cast_fp16_cast_int16)[name = string("transpose_7")];
+            tensor<fp16, [16, 224, 224]> _inversed_5149_cast_fp16 = mul(x = p2c_att_91_cast_fp16, y = _inversed_5149_y_0_to_fp16)[name = string("_inversed_5149_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_45_cast_fp16 = add(x = _inversed_5123_cast_fp16, y = _inversed_5149_cast_fp16)[name = string("rel_att_45_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_135_cast_fp16 = add(x = attention_scores_133_cast_fp16, y = rel_att_45_cast_fp16)[name = string("attention_scores_135_cast_fp16")];
+            tensor<int32, [4]> var_5154 = const()[name = string("op_5154"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_137_cast_fp16 = reshape(shape = var_5154, x = attention_scores_135_cast_fp16)[name = string("attention_scores_137_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_537_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_137_cast_fp16, cond = var_339)[name = string("input_537_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_539_cast_fp16 = softmax(axis = var_56, x = input_537_cast_fp16)[name = string("input_539_cast_fp16")];
+            tensor<int32, [3]> var_5163 = const()[name = string("op_5163"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_5164_cast_fp16 = reshape(shape = var_5163, x = input_539_cast_fp16)[name = string("op_5164_cast_fp16")];
             bool context_layer_89_transpose_x_0 = const()[name = string("context_layer_89_transpose_x_0"), val = bool(false)];
             bool context_layer_89_transpose_y_0 = const()[name = string("context_layer_89_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_89_cast_fp16 = matmul(transpose_x = context_layer_89_transpose_x_0, transpose_y = context_layer_89_transpose_y_0, x = var_5164_cast_fp16, y = value_layer_45_cast_fp16)[name = string("context_layer_89_cast_fp16")];
-            tensor<int32, [4]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5169_cast_fp16 = reshape(shape = var_5168, x = context_layer_89_cast_fp16)[name = string("op_5169_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_89_cast_fp16 = matmul(transpose_x = context_layer_89_transpose_x_0, transpose_y = context_layer_89_transpose_y_0, x = var_5164_cast_fp16, y = value_layer_45_cast_fp16)[name = string("context_layer_89_cast_fp16")];
+            tensor<int32, [4]> var_5168 = const()[name = string("op_5168"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5169_cast_fp16 = reshape(shape = var_5168, x = context_layer_89_cast_fp16)[name = string("op_5169_cast_fp16")];
             tensor<int32, [4]> var_5170 = const()[name = string("op_5170"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_5171_cast_fp16 = transpose(perm = var_5170, x = var_5169_cast_fp16)[name = string("transpose_6")];
-            tensor<fp16, [1, 512, 1024]> input_541_cast_fp16 = reshape(shape = var_5175, x = var_5171_cast_fp16)[name = string("input_541_cast_fp16")];
+            tensor<int32, [3]> var_5175 = const()[name = string("op_5175"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_5171_cast_fp16 = transpose(perm = var_5170, x = var_5169_cast_fp16)[name = string("transpose_6")];
+            tensor<fp16, [1, 224, 1024]> input_541_cast_fp16 = reshape(shape = var_5175, x = var_5171_cast_fp16)[name = string("input_541_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_22_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872011968)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874109184)))];
-            tensor<fp16, [1, 512, 1024]> linear_181_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_attention_output_dense_weight_to_fp16, x = input_541_cast_fp16)[name = string("linear_181_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_545_cast_fp16 = add(x = linear_181_cast_fp16, y = input_533_cast_fp16)[name = string("input_545_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_181_cast_fp16 = linear(bias = encoder_encoder_layer_22_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_attention_output_dense_weight_to_fp16, x = input_541_cast_fp16)[name = string("linear_181_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_545_cast_fp16 = add(x = linear_181_cast_fp16, y = input_533_cast_fp16)[name = string("input_545_cast_fp16")];
             tensor<int32, [1]> input_547_axes_0 = const()[name = string("input_547_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874111296)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874113408)))];
-            tensor<fp16, [1, 512, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_547_cast_fp16 = layer_norm(axes = input_547_axes_0, beta = encoder_encoder_layer_22_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_attention_output_LayerNorm_weight_to_fp16, x = input_545_cast_fp16)[name = string("input_547_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_22_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(874115520)))];
             tensor<fp16, [4096]> encoder_encoder_layer_22_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(882504192)))];
-            tensor<fp16, [1, 512, 4096]> linear_182_cast_fp16 = linear(bias = encoder_encoder_layer_22_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_22_intermediate_dense_weight_to_fp16, x = input_547_cast_fp16)[name = string("linear_182_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_182_cast_fp16 = linear(bias = encoder_encoder_layer_22_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_22_intermediate_dense_weight_to_fp16, x = input_547_cast_fp16)[name = string("linear_182_cast_fp16")];
             string input_551_mode_0 = const()[name = string("input_551_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_551_cast_fp16 = gelu(mode = input_551_mode_0, x = linear_182_cast_fp16)[name = string("input_551_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_551_cast_fp16 = gelu(mode = input_551_mode_0, x = linear_182_cast_fp16)[name = string("input_551_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_22_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(882512448)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890901120)))];
-            tensor<fp16, [1, 512, 1024]> linear_183_cast_fp16 = linear(bias = encoder_encoder_layer_22_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_output_dense_weight_to_fp16, x = input_551_cast_fp16)[name = string("linear_183_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_555_cast_fp16 = add(x = linear_183_cast_fp16, y = input_547_cast_fp16)[name = string("input_555_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_183_cast_fp16 = linear(bias = encoder_encoder_layer_22_output_dense_bias_to_fp16, weight = encoder_encoder_layer_22_output_dense_weight_to_fp16, x = input_551_cast_fp16)[name = string("linear_183_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_555_cast_fp16 = add(x = linear_183_cast_fp16, y = input_547_cast_fp16)[name = string("input_555_cast_fp16")];
             tensor<int32, [1]> input_557_axes_0 = const()[name = string("input_557_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890903232)))];
             tensor<fp16, [1024]> encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890905344)))];
-            tensor<fp16, [1, 512, 1024]> input_557_cast_fp16 = layer_norm(axes = input_557_axes_0, beta = encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16, x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_557_cast_fp16 = layer_norm(axes = input_557_axes_0, beta = encoder_encoder_layer_22_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_22_output_LayerNorm_weight_to_fp16, x = input_555_cast_fp16)[name = string("input_557_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(890907456)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893004672)))];
-            tensor<fp16, [1, 512, 1024]> linear_184_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_184_cast_fp16")];
-            tensor<int32, [4]> var_5226 = const()[name = string("op_5226"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_463_cast_fp16 = reshape(shape = var_5226, x = linear_184_cast_fp16)[name = string("x_463_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_184_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_query_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_query_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_184_cast_fp16")];
+            tensor<int32, [4]> var_5226 = const()[name = string("op_5226"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_463_cast_fp16 = reshape(shape = var_5226, x = linear_184_cast_fp16)[name = string("x_463_cast_fp16")];
             tensor<int32, [4]> var_5228 = const()[name = string("op_5228"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5229_cast_fp16 = transpose(perm = var_5228, x = x_463_cast_fp16)[name = string("transpose_5")];
-            tensor<fp16, [16, 512, 64]> query_layer_141_cast_fp16 = reshape(shape = var_5233, x = var_5229_cast_fp16)[name = string("query_layer_141_cast_fp16")];
+            tensor<int32, [3]> var_5233 = const()[name = string("op_5233"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5229_cast_fp16 = transpose(perm = var_5228, x = x_463_cast_fp16)[name = string("transpose_5")];
+            tensor<fp16, [16, 224, 64]> query_layer_141_cast_fp16 = reshape(shape = var_5233, x = var_5229_cast_fp16)[name = string("query_layer_141_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(893006784)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(895104000)))];
-            tensor<fp16, [1, 512, 1024]> linear_185_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_185_cast_fp16")];
-            tensor<int32, [4]> var_5240 = const()[name = string("op_5240"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_467_cast_fp16 = reshape(shape = var_5240, x = linear_185_cast_fp16)[name = string("x_467_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_185_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_key_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_key_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_185_cast_fp16")];
+            tensor<int32, [4]> var_5240 = const()[name = string("op_5240"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_467_cast_fp16 = reshape(shape = var_5240, x = linear_185_cast_fp16)[name = string("x_467_cast_fp16")];
             tensor<int32, [4]> var_5242 = const()[name = string("op_5242"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5247 = const()[name = string("op_5247"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5243_cast_fp16 = transpose(perm = var_5242, x = x_467_cast_fp16)[name = string("transpose_4")];
-            tensor<fp16, [16, 512, 64]> key_layer_cast_fp16 = reshape(shape = var_5247, x = var_5243_cast_fp16)[name = string("key_layer_cast_fp16")];
+            tensor<int32, [3]> var_5247 = const()[name = string("op_5247"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5243_cast_fp16 = transpose(perm = var_5242, x = x_467_cast_fp16)[name = string("transpose_4")];
+            tensor<fp16, [16, 224, 64]> key_layer_cast_fp16 = reshape(shape = var_5247, x = var_5243_cast_fp16)[name = string("key_layer_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(895106112)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897203328)))];
-            tensor<fp16, [1, 512, 1024]> linear_186_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_186_cast_fp16")];
-            tensor<int32, [4]> var_5254 = const()[name = string("op_5254"), val = tensor<int32, [4]>([1, 512, 16, -1])];
-            tensor<fp16, [1, 512, 16, 64]> x_471_cast_fp16 = reshape(shape = var_5254, x = linear_186_cast_fp16)[name = string("x_471_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_186_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_self_value_proj_bias_to_fp16, weight = encoder_encoder_layer_23_attention_self_value_proj_weight_to_fp16, x = input_557_cast_fp16)[name = string("linear_186_cast_fp16")];
+            tensor<int32, [4]> var_5254 = const()[name = string("op_5254"), val = tensor<int32, [4]>([1, 224, 16, -1])];
+            tensor<fp16, [1, 224, 16, 64]> x_471_cast_fp16 = reshape(shape = var_5254, x = linear_186_cast_fp16)[name = string("x_471_cast_fp16")];
             tensor<int32, [4]> var_5256 = const()[name = string("op_5256"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5261 = const()[name = string("op_5261"), val = tensor<int32, [3]>([-1, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5257_cast_fp16 = transpose(perm = var_5256, x = x_471_cast_fp16)[name = string("transpose_3")];
-            tensor<fp16, [16, 512, 64]> value_layer_cast_fp16 = reshape(shape = var_5261, x = var_5257_cast_fp16)[name = string("value_layer_cast_fp16")];
+            tensor<int32, [3]> var_5261 = const()[name = string("op_5261"), val = tensor<int32, [3]>([-1, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5257_cast_fp16 = transpose(perm = var_5256, x = x_471_cast_fp16)[name = string("transpose_3")];
+            tensor<fp16, [16, 224, 64]> value_layer_cast_fp16 = reshape(shape = var_5261, x = var_5257_cast_fp16)[name = string("value_layer_cast_fp16")];
             tensor<int32, [3]> var_5271_perm_0 = const()[name = string("op_5271_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5273_y_0_to_fp16 = const()[name = string("_inversed_5273_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 64, 512]> var_5271_cast_fp16 = transpose(perm = var_5271_perm_0, x = key_layer_cast_fp16)[name = string("transpose_2")];
-            tensor<fp16, [16, 64, 512]> _inversed_5273_cast_fp16 = mul(x = var_5271_cast_fp16, y = _inversed_5273_y_0_to_fp16)[name = string("_inversed_5273_cast_fp16")];
+            tensor<fp16, [16, 64, 224]> var_5271_cast_fp16 = transpose(perm = var_5271_perm_0, x = key_layer_cast_fp16)[name = string("transpose_2")];
+            tensor<fp16, [16, 64, 224]> _inversed_5273_cast_fp16 = mul(x = var_5271_cast_fp16, y = _inversed_5273_y_0_to_fp16)[name = string("_inversed_5273_cast_fp16")];
             bool attention_scores_139_transpose_x_0 = const()[name = string("attention_scores_139_transpose_x_0"), val = bool(false)];
             bool attention_scores_139_transpose_y_0 = const()[name = string("attention_scores_139_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> attention_scores_139_cast_fp16 = matmul(transpose_x = attention_scores_139_transpose_x_0, transpose_y = attention_scores_139_transpose_y_0, x = query_layer_141_cast_fp16, y = _inversed_5273_cast_fp16)[name = string("attention_scores_139_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_139_cast_fp16 = matmul(transpose_x = attention_scores_139_transpose_x_0, transpose_y = attention_scores_139_transpose_y_0, x = query_layer_141_cast_fp16, y = _inversed_5273_cast_fp16)[name = string("attention_scores_139_cast_fp16")];
             bool c2p_att_93_transpose_x_0 = const()[name = string("c2p_att_93_transpose_x_0"), val = bool(false)];
             bool c2p_att_93_transpose_y_0 = const()[name = string("c2p_att_93_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5329_to_fp16 = const()[name = string("op_5329_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897205440)))];
-            tensor<fp16, [16, 512, 512]> c2p_att_93_cast_fp16 = matmul(transpose_x = c2p_att_93_transpose_x_0, transpose_y = c2p_att_93_transpose_y_0, x = query_layer_141_cast_fp16, y = var_5329_to_fp16)[name = string("c2p_att_93_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> c2p_att_93_cast_fp16 = matmul(transpose_x = c2p_att_93_transpose_x_0, transpose_y = c2p_att_93_transpose_y_0, x = query_layer_141_cast_fp16, y = var_5329_to_fp16)[name = string("c2p_att_93_cast_fp16")];
             bool c2p_att_validate_indices_0 = const()[name = string("c2p_att_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> c2p_att_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_validate_indices_0, x = c2p_att_93_cast_fp16)[name = string("c2p_att_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> c2p_att_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_302_to_uint16, validate_indices = c2p_att_validate_indices_0, x = c2p_att_93_cast_fp16)[name = string("c2p_att_cast_fp16_cast_int16")];
             tensor<fp16, [1]> _inversed_5342_y_0_to_fp16 = const()[name = string("_inversed_5342_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> _inversed_5342_cast_fp16 = mul(x = c2p_att_cast_fp16_cast_int16, y = _inversed_5342_y_0_to_fp16)[name = string("_inversed_5342_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> _inversed_5342_cast_fp16 = mul(x = c2p_att_cast_fp16_cast_int16, y = _inversed_5342_y_0_to_fp16)[name = string("_inversed_5342_cast_fp16")];
             bool p2c_att_93_transpose_x_0 = const()[name = string("p2c_att_93_transpose_x_0"), val = bool(false)];
             bool p2c_att_93_transpose_y_0 = const()[name = string("p2c_att_93_transpose_y_0"), val = bool(false)];
             tensor<fp16, [16, 64, 512]> var_5357_to_fp16 = const()[name = string("op_5357_to_fp16"), val = tensor<fp16, [16, 64, 512]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(898254080)))];
-            tensor<fp16, [16, 512, 512]> p2c_att_93_cast_fp16 = matmul(transpose_x = p2c_att_93_transpose_x_0, transpose_y = p2c_att_93_transpose_y_0, x = key_layer_cast_fp16, y = var_5357_to_fp16)[name = string("p2c_att_93_cast_fp16")];
+            tensor<fp16, [16, 224, 512]> p2c_att_93_cast_fp16 = matmul(transpose_x = p2c_att_93_transpose_x_0, transpose_y = p2c_att_93_transpose_y_0, x = key_layer_cast_fp16, y = var_5357_to_fp16)[name = string("p2c_att_93_cast_fp16")];
             bool var_5365_validate_indices_0 = const()[name = string("op_5365_validate_indices_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 512]> var_5365_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5365_validate_indices_0, x = p2c_att_93_cast_fp16)[name = string("op_5365_cast_fp16_cast_int16")];
+            tensor<fp16, [16, 224, 224]> var_5365_cast_fp16_cast_int16 = gather_along_axis(axis = var_56, indices = var_327_to_uint16, validate_indices = var_5365_validate_indices_0, x = p2c_att_93_cast_fp16)[name = string("op_5365_cast_fp16_cast_int16")];
             tensor<int32, [3]> p2c_att_perm_0 = const()[name = string("p2c_att_perm_0"), val = tensor<int32, [3]>([0, -1, -2])];
             tensor<fp16, [1]> _inversed_5368_y_0_to_fp16 = const()[name = string("_inversed_5368_y_0_to_fp16"), val = tensor<fp16, [1]>([0x1.278p-4])];
-            tensor<fp16, [16, 512, 512]> p2c_att_cast_fp16 = transpose(perm = p2c_att_perm_0, x = var_5365_cast_fp16_cast_int16)[name = string("transpose_1")];
-            tensor<fp16, [16, 512, 512]> _inversed_5368_cast_fp16 = mul(x = p2c_att_cast_fp16, y = _inversed_5368_y_0_to_fp16)[name = string("_inversed_5368_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> rel_att_cast_fp16 = add(x = _inversed_5342_cast_fp16, y = _inversed_5368_cast_fp16)[name = string("rel_att_cast_fp16")];
-            tensor<fp16, [16, 512, 512]> attention_scores_141_cast_fp16 = add(x = attention_scores_139_cast_fp16, y = rel_att_cast_fp16)[name = string("attention_scores_141_cast_fp16")];
-            tensor<int32, [4]> var_5373 = const()[name = string("op_5373"), val = tensor<int32, [4]>([-1, 16, 512, 512])];
-            tensor<fp16, [1, 16, 512, 512]> attention_scores_cast_fp16 = reshape(shape = var_5373, x = attention_scores_141_cast_fp16)[name = string("attention_scores_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_561_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_cast_fp16, cond = var_339)[name = string("input_561_cast_fp16")];
-            tensor<fp16, [1, 16, 512, 512]> input_563_cast_fp16 = softmax(axis = var_56, x = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
-            tensor<int32, [3]> var_5382 = const()[name = string("op_5382"), val = tensor<int32, [3]>([-1, 512, 512])];
-            tensor<fp16, [16, 512, 512]> var_5383_cast_fp16 = reshape(shape = var_5382, x = input_563_cast_fp16)[name = string("op_5383_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> p2c_att_cast_fp16 = transpose(perm = p2c_att_perm_0, x = var_5365_cast_fp16_cast_int16)[name = string("transpose_1")];
+            tensor<fp16, [16, 224, 224]> _inversed_5368_cast_fp16 = mul(x = p2c_att_cast_fp16, y = _inversed_5368_y_0_to_fp16)[name = string("_inversed_5368_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> rel_att_cast_fp16 = add(x = _inversed_5342_cast_fp16, y = _inversed_5368_cast_fp16)[name = string("rel_att_cast_fp16")];
+            tensor<fp16, [16, 224, 224]> attention_scores_141_cast_fp16 = add(x = attention_scores_139_cast_fp16, y = rel_att_cast_fp16)[name = string("attention_scores_141_cast_fp16")];
+            tensor<int32, [4]> var_5373 = const()[name = string("op_5373"), val = tensor<int32, [4]>([-1, 16, 224, 224])];
+            tensor<fp16, [1, 16, 224, 224]> attention_scores_cast_fp16 = reshape(shape = var_5373, x = attention_scores_141_cast_fp16)[name = string("attention_scores_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_561_cast_fp16 = select(a = var_64_to_fp16, b = attention_scores_cast_fp16, cond = var_339)[name = string("input_561_cast_fp16")];
+            tensor<fp16, [1, 16, 224, 224]> input_563_cast_fp16 = softmax(axis = var_56, x = input_561_cast_fp16)[name = string("input_563_cast_fp16")];
+            tensor<int32, [3]> var_5382 = const()[name = string("op_5382"), val = tensor<int32, [3]>([-1, 224, 224])];
+            tensor<fp16, [16, 224, 224]> var_5383_cast_fp16 = reshape(shape = var_5382, x = input_563_cast_fp16)[name = string("op_5383_cast_fp16")];
             bool context_layer_93_transpose_x_0 = const()[name = string("context_layer_93_transpose_x_0"), val = bool(false)];
             bool context_layer_93_transpose_y_0 = const()[name = string("context_layer_93_transpose_y_0"), val = bool(false)];
-            tensor<fp16, [16, 512, 64]> context_layer_93_cast_fp16 = matmul(transpose_x = context_layer_93_transpose_x_0, transpose_y = context_layer_93_transpose_y_0, x = var_5383_cast_fp16, y = value_layer_cast_fp16)[name = string("context_layer_93_cast_fp16")];
-            tensor<int32, [4]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [4]>([-1, 16, 512, 64])];
-            tensor<fp16, [1, 16, 512, 64]> var_5388_cast_fp16 = reshape(shape = var_5387, x = context_layer_93_cast_fp16)[name = string("op_5388_cast_fp16")];
+            tensor<fp16, [16, 224, 64]> context_layer_93_cast_fp16 = matmul(transpose_x = context_layer_93_transpose_x_0, transpose_y = context_layer_93_transpose_y_0, x = var_5383_cast_fp16, y = value_layer_cast_fp16)[name = string("context_layer_93_cast_fp16")];
+            tensor<int32, [4]> var_5387 = const()[name = string("op_5387"), val = tensor<int32, [4]>([-1, 16, 224, 64])];
+            tensor<fp16, [1, 16, 224, 64]> var_5388_cast_fp16 = reshape(shape = var_5387, x = context_layer_93_cast_fp16)[name = string("op_5388_cast_fp16")];
             tensor<int32, [4]> var_5389 = const()[name = string("op_5389"), val = tensor<int32, [4]>([0, 2, 1, 3])];
-            tensor<int32, [3]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [3]>([1, 512, -1])];
-            tensor<fp16, [1, 512, 16, 64]> var_5390_cast_fp16 = transpose(perm = var_5389, x = var_5388_cast_fp16)[name = string("transpose_0")];
-            tensor<fp16, [1, 512, 1024]> input_565_cast_fp16 = reshape(shape = var_5394, x = var_5390_cast_fp16)[name = string("input_565_cast_fp16")];
+            tensor<int32, [3]> var_5394 = const()[name = string("op_5394"), val = tensor<int32, [3]>([1, 224, -1])];
+            tensor<fp16, [1, 224, 16, 64]> var_5390_cast_fp16 = transpose(perm = var_5389, x = var_5388_cast_fp16)[name = string("transpose_0")];
+            tensor<fp16, [1, 224, 1024]> input_565_cast_fp16 = reshape(shape = var_5394, x = var_5390_cast_fp16)[name = string("input_565_cast_fp16")];
             tensor<fp16, [1024, 1024]> encoder_encoder_layer_23_attention_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(899302720)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901399936)))];
-            tensor<fp16, [1, 512, 1024]> linear_189_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_attention_output_dense_weight_to_fp16, x = input_565_cast_fp16)[name = string("linear_189_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_569_cast_fp16 = add(x = linear_189_cast_fp16, y = input_557_cast_fp16)[name = string("input_569_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_189_cast_fp16 = linear(bias = encoder_encoder_layer_23_attention_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_attention_output_dense_weight_to_fp16, x = input_565_cast_fp16)[name = string("linear_189_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_569_cast_fp16 = add(x = linear_189_cast_fp16, y = input_557_cast_fp16)[name = string("input_569_cast_fp16")];
             tensor<int32, [1]> input_571_axes_0 = const()[name = string("input_571_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901402048)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901404160)))];
-            tensor<fp16, [1, 512, 1024]> input_571_cast_fp16 = layer_norm(axes = input_571_axes_0, beta = encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_571_cast_fp16 = layer_norm(axes = input_571_axes_0, beta = encoder_encoder_layer_23_attention_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_attention_output_LayerNorm_weight_to_fp16, x = input_569_cast_fp16)[name = string("input_571_cast_fp16")];
             tensor<fp16, [4096, 1024]> encoder_encoder_layer_23_intermediate_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_intermediate_dense_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901406272)))];
             tensor<fp16, [4096]> encoder_encoder_layer_23_intermediate_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_intermediate_dense_bias_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(909794944)))];
-            tensor<fp16, [1, 512, 4096]> linear_190_cast_fp16 = linear(bias = encoder_encoder_layer_23_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_23_intermediate_dense_weight_to_fp16, x = input_571_cast_fp16)[name = string("linear_190_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> linear_190_cast_fp16 = linear(bias = encoder_encoder_layer_23_intermediate_dense_bias_to_fp16, weight = encoder_encoder_layer_23_intermediate_dense_weight_to_fp16, x = input_571_cast_fp16)[name = string("linear_190_cast_fp16")];
             string input_575_mode_0 = const()[name = string("input_575_mode_0"), val = string("EXACT")];
-            tensor<fp16, [1, 512, 4096]> input_575_cast_fp16 = gelu(mode = input_575_mode_0, x = linear_190_cast_fp16)[name = string("input_575_cast_fp16")];
+            tensor<fp16, [1, 224, 4096]> input_575_cast_fp16 = gelu(mode = input_575_mode_0, x = linear_190_cast_fp16)[name = string("input_575_cast_fp16")];
             tensor<fp16, [1024, 4096]> encoder_encoder_layer_23_output_dense_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_dense_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(909803200)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_dense_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_dense_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918191872)))];
-            tensor<fp16, [1, 512, 1024]> linear_191_cast_fp16 = linear(bias = encoder_encoder_layer_23_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_output_dense_weight_to_fp16, x = input_575_cast_fp16)[name = string("linear_191_cast_fp16")];
-            tensor<fp16, [1, 512, 1024]> input_579_cast_fp16 = add(x = linear_191_cast_fp16, y = input_571_cast_fp16)[name = string("input_579_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> linear_191_cast_fp16 = linear(bias = encoder_encoder_layer_23_output_dense_bias_to_fp16, weight = encoder_encoder_layer_23_output_dense_weight_to_fp16, x = input_575_cast_fp16)[name = string("linear_191_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> input_579_cast_fp16 = add(x = linear_191_cast_fp16, y = input_571_cast_fp16)[name = string("input_579_cast_fp16")];
             tensor<int32, [1]> hidden_axes_0 = const()[name = string("hidden_axes_0"), val = tensor<int32, [1]>([-1])];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918193984)))];
             tensor<fp16, [1024]> encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16 = const()[name = string("encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918196096)))];
-            tensor<fp16, [1, 512, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, beta = encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16, x = input_579_cast_fp16)[name = string("hidden_cast_fp16")];
+            tensor<fp16, [1, 224, 1024]> hidden_cast_fp16 = layer_norm(axes = hidden_axes_0, beta = encoder_encoder_layer_23_output_LayerNorm_bias_to_fp16, epsilon = var_67_to_fp16, gamma = encoder_encoder_layer_23_output_LayerNorm_weight_to_fp16, x = input_579_cast_fp16)[name = string("hidden_cast_fp16")];
             tensor<int32, [3]> var_5426 = const()[name = string("op_5426"), val = tensor<int32, [3]>([1, 16, 1])];
             tensor<int32, [1, 16, 1]> var_5427 = reshape(shape = var_5426, x = label_positions)[name = string("op_5427")];
-            tensor<bool, [1, 16, 512]> var_5428 = equal(x = var_5427, y = slots)[name = string("op_5428")];
+            tensor<bool, [1, 16, 224]> var_5428 = equal(x = var_5427, y = slots)[name = string("op_5428")];
             bool input_581_transpose_x_0 = const()[name = string("input_581_transpose_x_0"), val = bool(false)];
             bool input_581_transpose_y_0 = const()[name = string("input_581_transpose_y_0"), val = bool(false)];
             string onehot_to_fp16_dtype_0 = const()[name = string("onehot_to_fp16_dtype_0"), val = string("fp16")];
-            tensor<fp16, [1, 16, 512]> var_5428_to_fp16 = cast(dtype = onehot_to_fp16_dtype_0, x = var_5428)[name = string("cast_490")];
+            tensor<fp16, [1, 16, 224]> var_5428_to_fp16 = cast(dtype = onehot_to_fp16_dtype_0, x = var_5428)[name = string("cast_490")];
             tensor<fp16, [1, 16, 1024]> input_581_cast_fp16 = matmul(transpose_x = input_581_transpose_x_0, transpose_y = input_581_transpose_y_0, x = var_5428_to_fp16, y = hidden_cast_fp16)[name = string("input_581_cast_fp16")];
             tensor<fp16, [2048, 1024]> classifier_0_weight_to_fp16 = const()[name = string("classifier_0_weight_to_fp16"), val = tensor<fp16, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(918198208)))];
             tensor<fp16, [2048]> classifier_0_bias_to_fp16 = const()[name = string("classifier_0_bias_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922392576)))];