program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3520.4.1"}, {"coremlc-version", "3520.5.1"}, {"coremltools-component-torch", "2.6.0"}, {"coremltools-source-dialect", "TorchScript"}, {"coremltools-version", "8.3.0"}})]
{
    func main<ios18>(tensor<fp16, [16, 1024]> block_cache_map, tensor<fp16, [1, 30720, 1, 1024]> key_cache, tensor<fp16, [1, 1024]> key_padding_mask, tensor<fp16, [1, 1, 1024]> shift_ctx, tensor<int32, [1, 16]> speech_pos_ids, tensor<int32, [1, 16]> speech_token_ids, tensor<fp16, [1, 30720, 1, 1024]> value_cache) {
            int32 speech_emb_1_axis_0 = const()[name = string("speech_emb_1_axis_0"), val = int32(0)];
            int32 speech_emb_1_batch_dims_0 = const()[name = string("speech_emb_1_batch_dims_0"), val = int32(0)];
            bool speech_emb_1_validate_indices_0 = const()[name = string("speech_emb_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [8195, 1024]> speech_emb_weight_to_fp16 = const()[name = string("speech_emb_weight_to_fp16"), val = tensor<fp16, [8195, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            string speech_token_ids_to_int16_dtype_0 = const()[name = string("speech_token_ids_to_int16_dtype_0"), val = string("int16")];
            tensor<int16, [1, 16]> speech_token_ids_to_int16 = cast(dtype = speech_token_ids_to_int16_dtype_0, x = speech_token_ids)[name = string("cast_534")];
            tensor<fp16, [1, 16, 1024]> speech_emb_1_cast_fp16_cast_uint16 = gather(axis = speech_emb_1_axis_0, batch_dims = speech_emb_1_batch_dims_0, indices = speech_token_ids_to_int16, validate_indices = speech_emb_1_validate_indices_0, x = speech_emb_weight_to_fp16)[name = string("speech_emb_1_cast_fp16_cast_uint16")];
            int32 var_790_axis_0 = const()[name = string("op_790_axis_0"), val = int32(0)];
            int32 var_790_batch_dims_0 = const()[name = string("op_790_batch_dims_0"), val = int32(0)];
            bool var_790_validate_indices_0 = const()[name = string("op_790_validate_indices_0"), val = bool(false)];
            tensor<fp16, [4100, 1024]> speech_pos_emb_emb_weight_to_fp16 = const()[name = string("speech_pos_emb_emb_weight_to_fp16"), val = tensor<fp16, [4100, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(16783488)))];
            string speech_pos_ids_to_uint16_dtype_0 = const()[name = string("speech_pos_ids_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1, 16]> speech_pos_ids_to_uint16 = cast(dtype = speech_pos_ids_to_uint16_dtype_0, x = speech_pos_ids)[name = string("cast_533")];
            tensor<fp16, [1, 16, 1024]> var_790_cast_fp16_cast_uint16 = gather(axis = var_790_axis_0, batch_dims = var_790_batch_dims_0, indices = speech_pos_ids_to_uint16, validate_indices = var_790_validate_indices_0, x = speech_pos_emb_emb_weight_to_fp16)[name = string("op_790_cast_fp16_cast_uint16")];
            tensor<fp16, [1, 16, 1024]> inputs_embeds_cast_fp16 = add(x = speech_emb_1_cast_fp16_cast_uint16, y = var_790_cast_fp16_cast_uint16)[name = string("inputs_embeds_cast_fp16")];
            tensor<int32, [1]> var_944_axes_0 = const()[name = string("op_944_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<int32, [1, 1, 16]> var_944 = expand_dims(axes = var_944_axes_0, x = speech_pos_ids)[name = string("op_944")];
            bool var_955_transpose_x_0 = const()[name = string("op_955_transpose_x_0"), val = bool(false)];
            bool var_955_transpose_y_0 = const()[name = string("op_955_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 32, 1]> const_2_to_fp16 = const()[name = string("const_2_to_fp16"), val = tensor<fp16, [1, 32, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25180352)))];
            string cast_11_to_fp16_dtype_0 = const()[name = string("cast_11_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [1, 1, 16]> var_944_to_fp16 = cast(dtype = cast_11_to_fp16_dtype_0, x = var_944)[name = string("cast_532")];
            tensor<fp16, [1, 32, 16]> var_955_cast_fp16 = matmul(transpose_x = var_955_transpose_x_0, transpose_y = var_955_transpose_y_0, x = const_2_to_fp16, y = var_944_to_fp16)[name = string("op_955_cast_fp16")];
            tensor<int32, [3]> freqs_perm_0 = const()[name = string("freqs_perm_0"), val = tensor<int32, [3]>([0, 2, 1])];
            int32 var_960 = const()[name = string("op_960"), val = int32(-1)];
            bool emb_1_interleave_0 = const()[name = string("emb_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 32]> freqs_cast_fp16 = transpose(perm = freqs_perm_0, x = var_955_cast_fp16)[name = string("transpose_180")];
            tensor<fp16, [1, 16, 64]> emb_1_cast_fp16 = concat(axis = var_960, interleave = emb_1_interleave_0, values = (freqs_cast_fp16, freqs_cast_fp16))[name = string("emb_1_cast_fp16")];
            tensor<fp16, [1, 16, 64]> var_962_cast_fp16 = cos(x = emb_1_cast_fp16)[name = string("op_962_cast_fp16")];
            tensor<fp16, [1, 16, 64]> var_970_cast_fp16 = sin(x = emb_1_cast_fp16)[name = string("op_970_cast_fp16")];
            tensor<int32, [1]> var_987_axes_0 = const()[name = string("op_987_axes_0"), val = tensor<int32, [1]>([0])];
            bool var_987_keep_dims_0 = const()[name = string("op_987_keep_dims_0"), val = bool(false)];
            tensor<fp16, [1024]> var_987_cast_fp16 = reduce_sum(axes = var_987_axes_0, keep_dims = var_987_keep_dims_0, x = block_cache_map)[name = string("op_987_cast_fp16")];
            tensor<int32, [4]> var_992 = const()[name = string("op_992"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<fp16, [1, 1, 1, 1024]> update_mask_cast_fp16 = reshape(shape = var_992, x = var_987_cast_fp16)[name = string("update_mask_cast_fp16")];
            tensor<int32, [4]> lkc_1_begin_0 = const()[name = string("lkc_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> lkc_1_end_0 = const()[name = string("lkc_1_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> lkc_1_end_mask_0 = const()[name = string("lkc_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_1_cast_fp16 = slice_by_index(begin = lkc_1_begin_0, end = lkc_1_end_0, end_mask = lkc_1_end_mask_0, x = key_cache)[name = string("lkc_1_cast_fp16")];
            tensor<int32, [4]> lvc_1_begin_0 = const()[name = string("lvc_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> lvc_1_end_0 = const()[name = string("lvc_1_end_0"), val = tensor<int32, [4]>([1, 1024, 1, 1024])];
            tensor<bool, [4]> lvc_1_end_mask_0 = const()[name = string("lvc_1_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_1_cast_fp16 = slice_by_index(begin = lvc_1_begin_0, end = lvc_1_end_0, end_mask = lvc_1_end_mask_0, x = value_cache)[name = string("lvc_1_cast_fp16")];
            fp16 var_1017_promoted_to_fp16 = const()[name = string("op_1017_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1023_cast_fp16 = pow(x = inputs_embeds_cast_fp16, y = var_1017_promoted_to_fp16)[name = string("op_1023_cast_fp16")];
            tensor<int32, [1]> variance_1_axes_0 = const()[name = string("variance_1_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_1_keep_dims_0 = const()[name = string("variance_1_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_1_cast_fp16 = reduce_mean(axes = variance_1_axes_0, keep_dims = variance_1_keep_dims_0, x = var_1023_cast_fp16)[name = string("variance_1_cast_fp16")];
            fp16 var_1026_to_fp16 = const()[name = string("op_1026_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1027_cast_fp16 = add(x = variance_1_cast_fp16, y = var_1026_to_fp16)[name = string("op_1027_cast_fp16")];
            fp32 var_1028_epsilon_0 = const()[name = string("op_1028_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1028_cast_fp16 = rsqrt(epsilon = var_1028_epsilon_0, x = var_1027_cast_fp16)[name = string("op_1028_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_3_cast_fp16 = mul(x = inputs_embeds_cast_fp16, y = var_1028_cast_fp16)[name = string("hidden_states_3_cast_fp16")];
            tensor<fp16, [1024]> layers_0_input_layernorm_weight_to_fp16 = const()[name = string("layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25180480)))];
            tensor<fp16, [1, 16, 1024]> hidden_1_cast_fp16 = mul(x = layers_0_input_layernorm_weight_to_fp16, y = hidden_states_3_cast_fp16)[name = string("hidden_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(25182592)))];
            tensor<fp16, [1024]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27279808)))];
            tensor<fp16, [1, 16, 1024]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_q_proj_weight_to_fp16, x = hidden_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_1052 = const()[name = string("op_1052"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_1_cast_fp16 = reshape(shape = var_1052, x = linear_0_cast_fp16)[name = string("q_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(27281920)))];
            tensor<fp16, [1, 16, 1024]> linear_1_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_k_proj_weight_to_fp16, x = hidden_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1059 = const()[name = string("op_1059"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_1_cast_fp16 = reshape(shape = var_1059, x = linear_1_cast_fp16)[name = string("k_1_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(29379136)))];
            tensor<fp16, [1, 16, 1024]> linear_2_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_v_proj_weight_to_fp16, x = hidden_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1066 = const()[name = string("op_1066"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_1_cast_fp16 = reshape(shape = var_1066, x = linear_2_cast_fp16)[name = string("v_1_cast_fp16")];
            tensor<int32, [4]> q_3_perm_0 = const()[name = string("q_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_3_perm_0 = const()[name = string("k_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_0_perm_0 = const()[name = string("transpose_0_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<int32, [1]> var_1078_axes_0 = const()[name = string("op_1078_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 16, 64]> var_1078_cast_fp16 = expand_dims(axes = var_1078_axes_0, x = var_962_cast_fp16)[name = string("op_1078_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_3_cast_fp16 = transpose(perm = q_3_perm_0, x = q_1_cast_fp16)[name = string("transpose_179")];
            tensor<fp16, [1, 16, 16, 64]> var_1079_cast_fp16 = mul(x = q_3_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1079_cast_fp16")];
            tensor<int32, [4]> x1_1_begin_0 = const()[name = string("x1_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_1_end_0 = const()[name = string("x1_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_1_end_mask_0 = const()[name = string("x1_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_1_cast_fp16 = slice_by_index(begin = x1_1_begin_0, end = x1_1_end_0, end_mask = x1_1_end_mask_0, x = q_3_cast_fp16)[name = string("x1_1_cast_fp16")];
            tensor<int32, [4]> x2_1_begin_0 = const()[name = string("x2_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_1_end_0 = const()[name = string("x2_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_1_end_mask_0 = const()[name = string("x2_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_1_cast_fp16 = slice_by_index(begin = x2_1_begin_0, end = x2_1_end_0, end_mask = x2_1_end_mask_0, x = q_3_cast_fp16)[name = string("x2_1_cast_fp16")];
            fp16 const_7_promoted_to_fp16 = const()[name = string("const_7_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1100_cast_fp16 = mul(x = x2_1_cast_fp16, y = const_7_promoted_to_fp16)[name = string("op_1100_cast_fp16")];
            int32 var_1102 = const()[name = string("op_1102"), val = int32(-1)];
            bool var_1103_interleave_0 = const()[name = string("op_1103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1103_cast_fp16 = concat(axis = var_1102, interleave = var_1103_interleave_0, values = (var_1100_cast_fp16, x1_1_cast_fp16))[name = string("op_1103_cast_fp16")];
            tensor<int32, [1]> var_1105_axes_0 = const()[name = string("op_1105_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 16, 64]> var_1105_cast_fp16 = expand_dims(axes = var_1105_axes_0, x = var_970_cast_fp16)[name = string("op_1105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1106_cast_fp16 = mul(x = var_1103_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1106_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_5_cast_fp16 = add(x = var_1079_cast_fp16, y = var_1106_cast_fp16)[name = string("q_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_3_cast_fp16 = transpose(perm = k_3_perm_0, x = k_1_cast_fp16)[name = string("transpose_178")];
            tensor<fp16, [1, 16, 16, 64]> var_1111_cast_fp16 = mul(x = k_3_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1111_cast_fp16")];
            tensor<int32, [4]> x1_3_begin_0 = const()[name = string("x1_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_3_end_0 = const()[name = string("x1_3_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_3_end_mask_0 = const()[name = string("x1_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_3_cast_fp16 = slice_by_index(begin = x1_3_begin_0, end = x1_3_end_0, end_mask = x1_3_end_mask_0, x = k_3_cast_fp16)[name = string("x1_3_cast_fp16")];
            tensor<int32, [4]> x2_3_begin_0 = const()[name = string("x2_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_3_end_0 = const()[name = string("x2_3_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_3_end_mask_0 = const()[name = string("x2_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_3_cast_fp16 = slice_by_index(begin = x2_3_begin_0, end = x2_3_end_0, end_mask = x2_3_end_mask_0, x = k_3_cast_fp16)[name = string("x2_3_cast_fp16")];
            fp16 const_10_promoted_to_fp16 = const()[name = string("const_10_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1132_cast_fp16 = mul(x = x2_3_cast_fp16, y = const_10_promoted_to_fp16)[name = string("op_1132_cast_fp16")];
            int32 var_1134 = const()[name = string("op_1134"), val = int32(-1)];
            bool var_1135_interleave_0 = const()[name = string("op_1135_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1135_cast_fp16 = concat(axis = var_1134, interleave = var_1135_interleave_0, values = (var_1132_cast_fp16, x1_3_cast_fp16))[name = string("op_1135_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1138_cast_fp16 = mul(x = var_1135_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1138_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_1_cast_fp16 = add(x = var_1111_cast_fp16, y = var_1138_cast_fp16)[name = string("x_1_cast_fp16")];
            tensor<int32, [4]> var_1145 = const()[name = string("op_1145"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1157 = const()[name = string("op_1157"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_1146_cast_fp16 = transpose(perm = var_1145, x = x_1_cast_fp16)[name = string("transpose_176")];
            tensor<fp16, [1, 1024, 1, 16]> var_1158_cast_fp16 = reshape(shape = var_1157, x = var_1146_cast_fp16)[name = string("op_1158_cast_fp16")];
            tensor<int32, [1]> var_1160_axes_0 = const()[name = string("op_1160_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1160_cast_fp16 = squeeze(axes = var_1160_axes_0, x = var_1158_cast_fp16)[name = string("op_1160_cast_fp16")];
            bool var_1161_transpose_x_0 = const()[name = string("op_1161_transpose_x_0"), val = bool(false)];
            bool var_1161_transpose_y_0 = const()[name = string("op_1161_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1161_cast_fp16 = matmul(transpose_x = var_1161_transpose_x_0, transpose_y = var_1161_transpose_y_0, x = var_1160_cast_fp16, y = block_cache_map)[name = string("op_1161_cast_fp16")];
            tensor<int32, [1]> nk_full_1_axes_0 = const()[name = string("nk_full_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_1_cast_fp16 = expand_dims(axes = nk_full_1_axes_0, x = var_1161_cast_fp16)[name = string("nk_full_1_cast_fp16")];
            tensor<int32, [4]> var_1180 = const()[name = string("op_1180"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_0_cast_fp16 = transpose(perm = transpose_0_perm_0, x = v_1_cast_fp16)[name = string("transpose_177")];
            tensor<fp16, [1, 1024, 1, 16]> var_1181_cast_fp16 = reshape(shape = var_1180, x = transpose_0_cast_fp16)[name = string("op_1181_cast_fp16")];
            tensor<int32, [1]> var_1183_axes_0 = const()[name = string("op_1183_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1183_cast_fp16 = squeeze(axes = var_1183_axes_0, x = var_1181_cast_fp16)[name = string("op_1183_cast_fp16")];
            bool var_1184_transpose_x_0 = const()[name = string("op_1184_transpose_x_0"), val = bool(false)];
            bool var_1184_transpose_y_0 = const()[name = string("op_1184_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1184_cast_fp16 = matmul(transpose_x = var_1184_transpose_x_0, transpose_y = var_1184_transpose_y_0, x = var_1183_cast_fp16, y = block_cache_map)[name = string("op_1184_cast_fp16")];
            tensor<int32, [1]> nv_full_1_axes_0 = const()[name = string("nv_full_1_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_1_cast_fp16 = expand_dims(axes = nv_full_1_axes_0, x = var_1184_cast_fp16)[name = string("nv_full_1_cast_fp16")];
            fp16 var_1187_to_fp16 = const()[name = string("op_1187_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [1, 1, 1, 1024]> var_1189_cast_fp16 = sub(x = var_1187_to_fp16, y = update_mask_cast_fp16)[name = string("op_1189_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1190_cast_fp16 = mul(x = lkc_1_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1190_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1191_cast_fp16 = mul(x = nk_full_1_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1191_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_3_cast_fp16 = add(x = var_1190_cast_fp16, y = var_1191_cast_fp16)[name = string("lkc_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1197_cast_fp16 = mul(x = lvc_1_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1197_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1198_cast_fp16 = mul(x = nv_full_1_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1198_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_3_cast_fp16 = add(x = var_1197_cast_fp16, y = var_1198_cast_fp16)[name = string("lvc_3_cast_fp16")];
            tensor<int32, [1]> var_1202_axes_0 = const()[name = string("op_1202_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1202_cast_fp16 = squeeze(axes = var_1202_axes_0, x = lkc_3_cast_fp16)[name = string("op_1202_cast_fp16")];
            tensor<int32, [4]> var_1206 = const()[name = string("op_1206"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_1_cast_fp16 = reshape(shape = var_1206, x = var_1202_cast_fp16)[name = string("kc_1_cast_fp16")];
            tensor<int32, [1]> var_1209_axes_0 = const()[name = string("op_1209_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1209_cast_fp16 = squeeze(axes = var_1209_axes_0, x = lvc_3_cast_fp16)[name = string("op_1209_cast_fp16")];
            tensor<int32, [4]> var_1213 = const()[name = string("op_1213"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_1_cast_fp16 = reshape(shape = var_1213, x = var_1209_cast_fp16)[name = string("vc_1_cast_fp16")];
            bool var_1215_transpose_x_0 = const()[name = string("op_1215_transpose_x_0"), val = bool(false)];
            bool var_1215_transpose_y_0 = const()[name = string("op_1215_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_1215_cast_fp16 = matmul(transpose_x = var_1215_transpose_x_0, transpose_y = var_1215_transpose_y_0, x = q_5_cast_fp16, y = kc_1_cast_fp16)[name = string("op_1215_cast_fp16")];
            fp16 _inversed_scores_1_y_0_to_fp16 = const()[name = string("_inversed_scores_1_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_1_cast_fp16 = mul(x = var_1215_cast_fp16, y = _inversed_scores_1_y_0_to_fp16)[name = string("_inversed_scores_1_cast_fp16")];
            tensor<int32, [1]> var_1224_axes_0 = const()[name = string("op_1224_axes_0"), val = tensor<int32, [1]>([1])];
            tensor<fp16, [1, 1, 1024]> var_1224_cast_fp16 = expand_dims(axes = var_1224_axes_0, x = key_padding_mask)[name = string("op_1224_cast_fp16")];
            tensor<int32, [1]> var_1226_axes_0 = const()[name = string("op_1226_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1, 1, 1024]> var_1226_cast_fp16 = expand_dims(axes = var_1226_axes_0, x = var_1224_cast_fp16)[name = string("op_1226_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_3_cast_fp16 = add(x = _inversed_scores_1_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_3_cast_fp16")];
            int32 var_1244 = const()[name = string("op_1244"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_1246_cast_fp16 = softmax(axis = var_1244, x = scores_3_cast_fp16)[name = string("op_1246_cast_fp16")];
            bool attn_out_1_transpose_x_1 = const()[name = string("attn_out_1_transpose_x_1"), val = bool(false)];
            bool attn_out_1_transpose_y_1 = const()[name = string("attn_out_1_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_1_cast_fp16 = matmul(transpose_x = attn_out_1_transpose_x_1, transpose_y = attn_out_1_transpose_y_1, x = var_1246_cast_fp16, y = vc_1_cast_fp16)[name = string("attn_out_1_cast_fp16")];
            tensor<int32, [4]> var_1258_perm_0 = const()[name = string("op_1258_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1260 = const()[name = string("op_1260"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_1258_cast_fp16 = transpose(perm = var_1258_perm_0, x = attn_out_1_cast_fp16)[name = string("transpose_175")];
            tensor<fp16, [1, 16, 1024]> input_5_cast_fp16 = reshape(shape = var_1260, x = var_1258_cast_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_0_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_0_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(31476352)))];
            tensor<fp16, [1, 16, 1024]> linear_3_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_self_attn_o_proj_weight_to_fp16, x = input_5_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_7_cast_fp16 = add(x = inputs_embeds_cast_fp16, y = linear_3_cast_fp16)[name = string("hidden_states_7_cast_fp16")];
            fp16 var_1270_promoted_to_fp16 = const()[name = string("op_1270_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1276_cast_fp16 = pow(x = hidden_states_7_cast_fp16, y = var_1270_promoted_to_fp16)[name = string("op_1276_cast_fp16")];
            tensor<int32, [1]> variance_3_axes_0 = const()[name = string("variance_3_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_3_keep_dims_0 = const()[name = string("variance_3_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_3_cast_fp16 = reduce_mean(axes = variance_3_axes_0, keep_dims = variance_3_keep_dims_0, x = var_1276_cast_fp16)[name = string("variance_3_cast_fp16")];
            fp16 var_1279_to_fp16 = const()[name = string("op_1279_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1280_cast_fp16 = add(x = variance_3_cast_fp16, y = var_1279_to_fp16)[name = string("op_1280_cast_fp16")];
            fp32 var_1281_epsilon_0 = const()[name = string("op_1281_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1281_cast_fp16 = rsqrt(epsilon = var_1281_epsilon_0, x = var_1280_cast_fp16)[name = string("op_1281_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_11_cast_fp16 = mul(x = hidden_states_7_cast_fp16, y = var_1281_cast_fp16)[name = string("hidden_states_11_cast_fp16")];
            tensor<fp16, [1024]> layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33573568)))];
            tensor<fp16, [1, 16, 1024]> input_7_cast_fp16 = mul(x = layers_0_post_attention_layernorm_weight_to_fp16, y = hidden_states_11_cast_fp16)[name = string("input_7_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_0_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(33575680)))];
            tensor<fp16, [4096]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41964352)))];
            tensor<fp16, [1, 16, 4096]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_gate_proj_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_1294_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1294_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_0_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41972608)))];
            tensor<fp16, [1, 16, 4096]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_0_mlp_up_proj_weight_to_fp16, x = input_7_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_11_cast_fp16 = mul(x = var_1294_cast_fp16, y = linear_5_cast_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_0_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_0_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50361280)))];
            tensor<fp16, [1, 16, 1024]> linear_6_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_0_mlp_down_proj_weight_to_fp16, x = input_11_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_15_cast_fp16 = add(x = hidden_states_7_cast_fp16, y = linear_6_cast_fp16)[name = string("hidden_states_15_cast_fp16")];
            tensor<int32, [4]> lkc_5_begin_0 = const()[name = string("lkc_5_begin_0"), val = tensor<int32, [4]>([0, 1024, 0, 0])];
            tensor<int32, [4]> lkc_5_end_0 = const()[name = string("lkc_5_end_0"), val = tensor<int32, [4]>([1, 2048, 1, 1024])];
            tensor<bool, [4]> lkc_5_end_mask_0 = const()[name = string("lkc_5_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_5_cast_fp16 = slice_by_index(begin = lkc_5_begin_0, end = lkc_5_end_0, end_mask = lkc_5_end_mask_0, x = key_cache)[name = string("lkc_5_cast_fp16")];
            tensor<int32, [4]> lvc_5_begin_0 = const()[name = string("lvc_5_begin_0"), val = tensor<int32, [4]>([0, 1024, 0, 0])];
            tensor<int32, [4]> lvc_5_end_0 = const()[name = string("lvc_5_end_0"), val = tensor<int32, [4]>([1, 2048, 1, 1024])];
            tensor<bool, [4]> lvc_5_end_mask_0 = const()[name = string("lvc_5_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_5_cast_fp16 = slice_by_index(begin = lvc_5_begin_0, end = lvc_5_end_0, end_mask = lvc_5_end_mask_0, x = value_cache)[name = string("lvc_5_cast_fp16")];
            fp16 var_1325_promoted_to_fp16 = const()[name = string("op_1325_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1331_cast_fp16 = pow(x = hidden_states_15_cast_fp16, y = var_1325_promoted_to_fp16)[name = string("op_1331_cast_fp16")];
            tensor<int32, [1]> variance_5_axes_0 = const()[name = string("variance_5_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_5_keep_dims_0 = const()[name = string("variance_5_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_5_cast_fp16 = reduce_mean(axes = variance_5_axes_0, keep_dims = variance_5_keep_dims_0, x = var_1331_cast_fp16)[name = string("variance_5_cast_fp16")];
            fp16 var_1334_to_fp16 = const()[name = string("op_1334_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1335_cast_fp16 = add(x = variance_5_cast_fp16, y = var_1334_to_fp16)[name = string("op_1335_cast_fp16")];
            fp32 var_1336_epsilon_0 = const()[name = string("op_1336_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1336_cast_fp16 = rsqrt(epsilon = var_1336_epsilon_0, x = var_1335_cast_fp16)[name = string("op_1336_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_19_cast_fp16 = mul(x = hidden_states_15_cast_fp16, y = var_1336_cast_fp16)[name = string("hidden_states_19_cast_fp16")];
            tensor<fp16, [1024]> layers_1_input_layernorm_weight_to_fp16 = const()[name = string("layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58749952)))];
            tensor<fp16, [1, 16, 1024]> hidden_5_cast_fp16 = mul(x = layers_1_input_layernorm_weight_to_fp16, y = hidden_states_19_cast_fp16)[name = string("hidden_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(58752064)))];
            tensor<fp16, [1, 16, 1024]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_q_proj_weight_to_fp16, x = hidden_5_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_1360 = const()[name = string("op_1360"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_7_cast_fp16 = reshape(shape = var_1360, x = linear_7_cast_fp16)[name = string("q_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(60849280)))];
            tensor<fp16, [1, 16, 1024]> linear_8_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_k_proj_weight_to_fp16, x = hidden_5_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1367 = const()[name = string("op_1367"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_5_cast_fp16 = reshape(shape = var_1367, x = linear_8_cast_fp16)[name = string("k_5_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62946496)))];
            tensor<fp16, [1, 16, 1024]> linear_9_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_v_proj_weight_to_fp16, x = hidden_5_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1374 = const()[name = string("op_1374"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_3_cast_fp16 = reshape(shape = var_1374, x = linear_9_cast_fp16)[name = string("v_3_cast_fp16")];
            tensor<int32, [4]> q_9_perm_0 = const()[name = string("q_9_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_7_perm_0 = const()[name = string("k_7_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_1_perm_0 = const()[name = string("transpose_1_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_9_cast_fp16 = transpose(perm = q_9_perm_0, x = q_7_cast_fp16)[name = string("transpose_174")];
            tensor<fp16, [1, 16, 16, 64]> var_1387_cast_fp16 = mul(x = q_9_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1387_cast_fp16")];
            tensor<int32, [4]> x1_5_begin_0 = const()[name = string("x1_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_5_end_0 = const()[name = string("x1_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_5_end_mask_0 = const()[name = string("x1_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_5_cast_fp16 = slice_by_index(begin = x1_5_begin_0, end = x1_5_end_0, end_mask = x1_5_end_mask_0, x = q_9_cast_fp16)[name = string("x1_5_cast_fp16")];
            tensor<int32, [4]> x2_5_begin_0 = const()[name = string("x2_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_5_end_0 = const()[name = string("x2_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_5_end_mask_0 = const()[name = string("x2_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_5_cast_fp16 = slice_by_index(begin = x2_5_begin_0, end = x2_5_end_0, end_mask = x2_5_end_mask_0, x = q_9_cast_fp16)[name = string("x2_5_cast_fp16")];
            fp16 const_19_promoted_to_fp16 = const()[name = string("const_19_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1408_cast_fp16 = mul(x = x2_5_cast_fp16, y = const_19_promoted_to_fp16)[name = string("op_1408_cast_fp16")];
            int32 var_1410 = const()[name = string("op_1410"), val = int32(-1)];
            bool var_1411_interleave_0 = const()[name = string("op_1411_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1411_cast_fp16 = concat(axis = var_1410, interleave = var_1411_interleave_0, values = (var_1408_cast_fp16, x1_5_cast_fp16))[name = string("op_1411_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1414_cast_fp16 = mul(x = var_1411_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1414_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_11_cast_fp16 = add(x = var_1387_cast_fp16, y = var_1414_cast_fp16)[name = string("q_11_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_7_cast_fp16 = transpose(perm = k_7_perm_0, x = k_5_cast_fp16)[name = string("transpose_173")];
            tensor<fp16, [1, 16, 16, 64]> var_1419_cast_fp16 = mul(x = k_7_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1419_cast_fp16")];
            tensor<int32, [4]> x1_7_begin_0 = const()[name = string("x1_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_7_end_0 = const()[name = string("x1_7_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_7_end_mask_0 = const()[name = string("x1_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_7_cast_fp16 = slice_by_index(begin = x1_7_begin_0, end = x1_7_end_0, end_mask = x1_7_end_mask_0, x = k_7_cast_fp16)[name = string("x1_7_cast_fp16")];
            tensor<int32, [4]> x2_7_begin_0 = const()[name = string("x2_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_7_end_0 = const()[name = string("x2_7_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_7_end_mask_0 = const()[name = string("x2_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_7_cast_fp16 = slice_by_index(begin = x2_7_begin_0, end = x2_7_end_0, end_mask = x2_7_end_mask_0, x = k_7_cast_fp16)[name = string("x2_7_cast_fp16")];
            fp16 const_22_promoted_to_fp16 = const()[name = string("const_22_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1440_cast_fp16 = mul(x = x2_7_cast_fp16, y = const_22_promoted_to_fp16)[name = string("op_1440_cast_fp16")];
            int32 var_1442 = const()[name = string("op_1442"), val = int32(-1)];
            bool var_1443_interleave_0 = const()[name = string("op_1443_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1443_cast_fp16 = concat(axis = var_1442, interleave = var_1443_interleave_0, values = (var_1440_cast_fp16, x1_7_cast_fp16))[name = string("op_1443_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1446_cast_fp16 = mul(x = var_1443_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1446_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_5_cast_fp16 = add(x = var_1419_cast_fp16, y = var_1446_cast_fp16)[name = string("x_5_cast_fp16")];
            tensor<int32, [4]> var_1453 = const()[name = string("op_1453"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1465 = const()[name = string("op_1465"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_1454_cast_fp16 = transpose(perm = var_1453, x = x_5_cast_fp16)[name = string("transpose_171")];
            tensor<fp16, [1, 1024, 1, 16]> var_1466_cast_fp16 = reshape(shape = var_1465, x = var_1454_cast_fp16)[name = string("op_1466_cast_fp16")];
            tensor<int32, [1]> var_1468_axes_0 = const()[name = string("op_1468_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1468_cast_fp16 = squeeze(axes = var_1468_axes_0, x = var_1466_cast_fp16)[name = string("op_1468_cast_fp16")];
            bool var_1469_transpose_x_0 = const()[name = string("op_1469_transpose_x_0"), val = bool(false)];
            bool var_1469_transpose_y_0 = const()[name = string("op_1469_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1469_cast_fp16 = matmul(transpose_x = var_1469_transpose_x_0, transpose_y = var_1469_transpose_y_0, x = var_1468_cast_fp16, y = block_cache_map)[name = string("op_1469_cast_fp16")];
            tensor<int32, [1]> nk_full_3_axes_0 = const()[name = string("nk_full_3_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_3_cast_fp16 = expand_dims(axes = nk_full_3_axes_0, x = var_1469_cast_fp16)[name = string("nk_full_3_cast_fp16")];
            tensor<int32, [4]> var_1488 = const()[name = string("op_1488"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_1_cast_fp16 = transpose(perm = transpose_1_perm_0, x = v_3_cast_fp16)[name = string("transpose_172")];
            tensor<fp16, [1, 1024, 1, 16]> var_1489_cast_fp16 = reshape(shape = var_1488, x = transpose_1_cast_fp16)[name = string("op_1489_cast_fp16")];
            tensor<int32, [1]> var_1491_axes_0 = const()[name = string("op_1491_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1491_cast_fp16 = squeeze(axes = var_1491_axes_0, x = var_1489_cast_fp16)[name = string("op_1491_cast_fp16")];
            bool var_1492_transpose_x_0 = const()[name = string("op_1492_transpose_x_0"), val = bool(false)];
            bool var_1492_transpose_y_0 = const()[name = string("op_1492_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1492_cast_fp16 = matmul(transpose_x = var_1492_transpose_x_0, transpose_y = var_1492_transpose_y_0, x = var_1491_cast_fp16, y = block_cache_map)[name = string("op_1492_cast_fp16")];
            tensor<int32, [1]> nv_full_3_axes_0 = const()[name = string("nv_full_3_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_3_cast_fp16 = expand_dims(axes = nv_full_3_axes_0, x = var_1492_cast_fp16)[name = string("nv_full_3_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1498_cast_fp16 = mul(x = lkc_5_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1498_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1499_cast_fp16 = mul(x = nk_full_3_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1499_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_7_cast_fp16 = add(x = var_1498_cast_fp16, y = var_1499_cast_fp16)[name = string("lkc_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1505_cast_fp16 = mul(x = lvc_5_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1505_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1506_cast_fp16 = mul(x = nv_full_3_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1506_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_7_cast_fp16 = add(x = var_1505_cast_fp16, y = var_1506_cast_fp16)[name = string("lvc_7_cast_fp16")];
            tensor<int32, [1]> var_1510_axes_0 = const()[name = string("op_1510_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1510_cast_fp16 = squeeze(axes = var_1510_axes_0, x = lkc_7_cast_fp16)[name = string("op_1510_cast_fp16")];
            tensor<int32, [4]> var_1514 = const()[name = string("op_1514"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_3_cast_fp16 = reshape(shape = var_1514, x = var_1510_cast_fp16)[name = string("kc_3_cast_fp16")];
            tensor<int32, [1]> var_1517_axes_0 = const()[name = string("op_1517_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1517_cast_fp16 = squeeze(axes = var_1517_axes_0, x = lvc_7_cast_fp16)[name = string("op_1517_cast_fp16")];
            tensor<int32, [4]> var_1521 = const()[name = string("op_1521"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_3_cast_fp16 = reshape(shape = var_1521, x = var_1517_cast_fp16)[name = string("vc_3_cast_fp16")];
            bool var_1523_transpose_x_0 = const()[name = string("op_1523_transpose_x_0"), val = bool(false)];
            bool var_1523_transpose_y_0 = const()[name = string("op_1523_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_1523_cast_fp16 = matmul(transpose_x = var_1523_transpose_x_0, transpose_y = var_1523_transpose_y_0, x = q_11_cast_fp16, y = kc_3_cast_fp16)[name = string("op_1523_cast_fp16")];
            fp16 _inversed_scores_7_y_0_to_fp16 = const()[name = string("_inversed_scores_7_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_7_cast_fp16 = mul(x = var_1523_cast_fp16, y = _inversed_scores_7_y_0_to_fp16)[name = string("_inversed_scores_7_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_9_cast_fp16 = add(x = _inversed_scores_7_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_9_cast_fp16")];
            int32 var_1552 = const()[name = string("op_1552"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_1554_cast_fp16 = softmax(axis = var_1552, x = scores_9_cast_fp16)[name = string("op_1554_cast_fp16")];
            bool attn_out_5_transpose_x_1 = const()[name = string("attn_out_5_transpose_x_1"), val = bool(false)];
            bool attn_out_5_transpose_y_1 = const()[name = string("attn_out_5_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_5_cast_fp16 = matmul(transpose_x = attn_out_5_transpose_x_1, transpose_y = attn_out_5_transpose_y_1, x = var_1554_cast_fp16, y = vc_3_cast_fp16)[name = string("attn_out_5_cast_fp16")];
            tensor<int32, [4]> var_1566_perm_0 = const()[name = string("op_1566_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1568 = const()[name = string("op_1568"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_1566_cast_fp16 = transpose(perm = var_1566_perm_0, x = attn_out_5_cast_fp16)[name = string("transpose_170")];
            tensor<fp16, [1, 16, 1024]> input_13_cast_fp16 = reshape(shape = var_1568, x = var_1566_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_1_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_1_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(65043712)))];
            tensor<fp16, [1, 16, 1024]> linear_10_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_self_attn_o_proj_weight_to_fp16, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_23_cast_fp16 = add(x = hidden_states_15_cast_fp16, y = linear_10_cast_fp16)[name = string("hidden_states_23_cast_fp16")];
            fp16 var_1578_promoted_to_fp16 = const()[name = string("op_1578_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1584_cast_fp16 = pow(x = hidden_states_23_cast_fp16, y = var_1578_promoted_to_fp16)[name = string("op_1584_cast_fp16")];
            tensor<int32, [1]> variance_7_axes_0 = const()[name = string("variance_7_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_7_keep_dims_0 = const()[name = string("variance_7_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_7_cast_fp16 = reduce_mean(axes = variance_7_axes_0, keep_dims = variance_7_keep_dims_0, x = var_1584_cast_fp16)[name = string("variance_7_cast_fp16")];
            fp16 var_1587_to_fp16 = const()[name = string("op_1587_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1588_cast_fp16 = add(x = variance_7_cast_fp16, y = var_1587_to_fp16)[name = string("op_1588_cast_fp16")];
            fp32 var_1589_epsilon_0 = const()[name = string("op_1589_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1589_cast_fp16 = rsqrt(epsilon = var_1589_epsilon_0, x = var_1588_cast_fp16)[name = string("op_1589_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_27_cast_fp16 = mul(x = hidden_states_23_cast_fp16, y = var_1589_cast_fp16)[name = string("hidden_states_27_cast_fp16")];
            tensor<fp16, [1024]> layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67140928)))];
            tensor<fp16, [1, 16, 1024]> input_15_cast_fp16 = mul(x = layers_1_post_attention_layernorm_weight_to_fp16, y = hidden_states_27_cast_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_1_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67143040)))];
            tensor<fp16, [1, 16, 4096]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_gate_proj_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_1602_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_1602_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_1_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(75531712)))];
            tensor<fp16, [1, 16, 4096]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_1_mlp_up_proj_weight_to_fp16, x = input_15_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_19_cast_fp16 = mul(x = var_1602_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_1_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_1_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(83920384)))];
            tensor<fp16, [1, 16, 1024]> linear_13_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_1_mlp_down_proj_weight_to_fp16, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_31_cast_fp16 = add(x = hidden_states_23_cast_fp16, y = linear_13_cast_fp16)[name = string("hidden_states_31_cast_fp16")];
            tensor<int32, [4]> lkc_9_begin_0 = const()[name = string("lkc_9_begin_0"), val = tensor<int32, [4]>([0, 2048, 0, 0])];
            tensor<int32, [4]> lkc_9_end_0 = const()[name = string("lkc_9_end_0"), val = tensor<int32, [4]>([1, 3072, 1, 1024])];
            tensor<bool, [4]> lkc_9_end_mask_0 = const()[name = string("lkc_9_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_9_cast_fp16 = slice_by_index(begin = lkc_9_begin_0, end = lkc_9_end_0, end_mask = lkc_9_end_mask_0, x = key_cache)[name = string("lkc_9_cast_fp16")];
            tensor<int32, [4]> lvc_9_begin_0 = const()[name = string("lvc_9_begin_0"), val = tensor<int32, [4]>([0, 2048, 0, 0])];
            tensor<int32, [4]> lvc_9_end_0 = const()[name = string("lvc_9_end_0"), val = tensor<int32, [4]>([1, 3072, 1, 1024])];
            tensor<bool, [4]> lvc_9_end_mask_0 = const()[name = string("lvc_9_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_9_cast_fp16 = slice_by_index(begin = lvc_9_begin_0, end = lvc_9_end_0, end_mask = lvc_9_end_mask_0, x = value_cache)[name = string("lvc_9_cast_fp16")];
            fp16 var_1633_promoted_to_fp16 = const()[name = string("op_1633_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1639_cast_fp16 = pow(x = hidden_states_31_cast_fp16, y = var_1633_promoted_to_fp16)[name = string("op_1639_cast_fp16")];
            tensor<int32, [1]> variance_9_axes_0 = const()[name = string("variance_9_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_9_keep_dims_0 = const()[name = string("variance_9_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_9_cast_fp16 = reduce_mean(axes = variance_9_axes_0, keep_dims = variance_9_keep_dims_0, x = var_1639_cast_fp16)[name = string("variance_9_cast_fp16")];
            fp16 var_1642_to_fp16 = const()[name = string("op_1642_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1643_cast_fp16 = add(x = variance_9_cast_fp16, y = var_1642_to_fp16)[name = string("op_1643_cast_fp16")];
            fp32 var_1644_epsilon_0 = const()[name = string("op_1644_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1644_cast_fp16 = rsqrt(epsilon = var_1644_epsilon_0, x = var_1643_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_35_cast_fp16 = mul(x = hidden_states_31_cast_fp16, y = var_1644_cast_fp16)[name = string("hidden_states_35_cast_fp16")];
            tensor<fp16, [1024]> layers_2_input_layernorm_weight_to_fp16 = const()[name = string("layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92309056)))];
            tensor<fp16, [1, 16, 1024]> hidden_9_cast_fp16 = mul(x = layers_2_input_layernorm_weight_to_fp16, y = hidden_states_35_cast_fp16)[name = string("hidden_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92311168)))];
            tensor<fp16, [1, 16, 1024]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_q_proj_weight_to_fp16, x = hidden_9_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_1668 = const()[name = string("op_1668"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_13_cast_fp16 = reshape(shape = var_1668, x = linear_14_cast_fp16)[name = string("q_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(94408384)))];
            tensor<fp16, [1, 16, 1024]> linear_15_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_k_proj_weight_to_fp16, x = hidden_9_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_1675 = const()[name = string("op_1675"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_9_cast_fp16 = reshape(shape = var_1675, x = linear_15_cast_fp16)[name = string("k_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96505600)))];
            tensor<fp16, [1, 16, 1024]> linear_16_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_v_proj_weight_to_fp16, x = hidden_9_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_1682 = const()[name = string("op_1682"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_5_cast_fp16 = reshape(shape = var_1682, x = linear_16_cast_fp16)[name = string("v_5_cast_fp16")];
            tensor<int32, [4]> q_15_perm_0 = const()[name = string("q_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_11_perm_0 = const()[name = string("k_11_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_2_perm_0 = const()[name = string("transpose_2_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_15_cast_fp16 = transpose(perm = q_15_perm_0, x = q_13_cast_fp16)[name = string("transpose_169")];
            tensor<fp16, [1, 16, 16, 64]> var_1695_cast_fp16 = mul(x = q_15_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1695_cast_fp16")];
            tensor<int32, [4]> x1_9_begin_0 = const()[name = string("x1_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_9_end_0 = const()[name = string("x1_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_9_end_mask_0 = const()[name = string("x1_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_9_cast_fp16 = slice_by_index(begin = x1_9_begin_0, end = x1_9_end_0, end_mask = x1_9_end_mask_0, x = q_15_cast_fp16)[name = string("x1_9_cast_fp16")];
            tensor<int32, [4]> x2_9_begin_0 = const()[name = string("x2_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_9_end_0 = const()[name = string("x2_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_9_end_mask_0 = const()[name = string("x2_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_9_cast_fp16 = slice_by_index(begin = x2_9_begin_0, end = x2_9_end_0, end_mask = x2_9_end_mask_0, x = q_15_cast_fp16)[name = string("x2_9_cast_fp16")];
            fp16 const_31_promoted_to_fp16 = const()[name = string("const_31_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1716_cast_fp16 = mul(x = x2_9_cast_fp16, y = const_31_promoted_to_fp16)[name = string("op_1716_cast_fp16")];
            int32 var_1718 = const()[name = string("op_1718"), val = int32(-1)];
            bool var_1719_interleave_0 = const()[name = string("op_1719_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1719_cast_fp16 = concat(axis = var_1718, interleave = var_1719_interleave_0, values = (var_1716_cast_fp16, x1_9_cast_fp16))[name = string("op_1719_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1722_cast_fp16 = mul(x = var_1719_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1722_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_17_cast_fp16 = add(x = var_1695_cast_fp16, y = var_1722_cast_fp16)[name = string("q_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_11_cast_fp16 = transpose(perm = k_11_perm_0, x = k_9_cast_fp16)[name = string("transpose_168")];
            tensor<fp16, [1, 16, 16, 64]> var_1727_cast_fp16 = mul(x = k_11_cast_fp16, y = var_1078_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<int32, [4]> x1_11_begin_0 = const()[name = string("x1_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_11_end_0 = const()[name = string("x1_11_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_11_end_mask_0 = const()[name = string("x1_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_11_cast_fp16 = slice_by_index(begin = x1_11_begin_0, end = x1_11_end_0, end_mask = x1_11_end_mask_0, x = k_11_cast_fp16)[name = string("x1_11_cast_fp16")];
            tensor<int32, [4]> x2_11_begin_0 = const()[name = string("x2_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_11_end_0 = const()[name = string("x2_11_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_11_end_mask_0 = const()[name = string("x2_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_11_cast_fp16 = slice_by_index(begin = x2_11_begin_0, end = x2_11_end_0, end_mask = x2_11_end_mask_0, x = k_11_cast_fp16)[name = string("x2_11_cast_fp16")];
            fp16 const_34_promoted_to_fp16 = const()[name = string("const_34_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_1748_cast_fp16 = mul(x = x2_11_cast_fp16, y = const_34_promoted_to_fp16)[name = string("op_1748_cast_fp16")];
            int32 var_1750 = const()[name = string("op_1750"), val = int32(-1)];
            bool var_1751_interleave_0 = const()[name = string("op_1751_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_1751_cast_fp16 = concat(axis = var_1750, interleave = var_1751_interleave_0, values = (var_1748_cast_fp16, x1_11_cast_fp16))[name = string("op_1751_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1754_cast_fp16 = mul(x = var_1751_cast_fp16, y = var_1105_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_9_cast_fp16 = add(x = var_1727_cast_fp16, y = var_1754_cast_fp16)[name = string("x_9_cast_fp16")];
            tensor<int32, [4]> var_1761 = const()[name = string("op_1761"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_1773 = const()[name = string("op_1773"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_1762_cast_fp16 = transpose(perm = var_1761, x = x_9_cast_fp16)[name = string("transpose_166")];
            tensor<fp16, [1, 1024, 1, 16]> var_1774_cast_fp16 = reshape(shape = var_1773, x = var_1762_cast_fp16)[name = string("op_1774_cast_fp16")];
            tensor<int32, [1]> var_1776_axes_0 = const()[name = string("op_1776_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1776_cast_fp16 = squeeze(axes = var_1776_axes_0, x = var_1774_cast_fp16)[name = string("op_1776_cast_fp16")];
            bool var_1777_transpose_x_0 = const()[name = string("op_1777_transpose_x_0"), val = bool(false)];
            bool var_1777_transpose_y_0 = const()[name = string("op_1777_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1777_cast_fp16 = matmul(transpose_x = var_1777_transpose_x_0, transpose_y = var_1777_transpose_y_0, x = var_1776_cast_fp16, y = block_cache_map)[name = string("op_1777_cast_fp16")];
            tensor<int32, [1]> nk_full_5_axes_0 = const()[name = string("nk_full_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_5_cast_fp16 = expand_dims(axes = nk_full_5_axes_0, x = var_1777_cast_fp16)[name = string("nk_full_5_cast_fp16")];
            tensor<int32, [4]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_2_cast_fp16 = transpose(perm = transpose_2_perm_0, x = v_5_cast_fp16)[name = string("transpose_167")];
            tensor<fp16, [1, 1024, 1, 16]> var_1797_cast_fp16 = reshape(shape = var_1796, x = transpose_2_cast_fp16)[name = string("op_1797_cast_fp16")];
            tensor<int32, [1]> var_1799_axes_0 = const()[name = string("op_1799_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_1799_cast_fp16 = squeeze(axes = var_1799_axes_0, x = var_1797_cast_fp16)[name = string("op_1799_cast_fp16")];
            bool var_1800_transpose_x_0 = const()[name = string("op_1800_transpose_x_0"), val = bool(false)];
            bool var_1800_transpose_y_0 = const()[name = string("op_1800_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_1800_cast_fp16 = matmul(transpose_x = var_1800_transpose_x_0, transpose_y = var_1800_transpose_y_0, x = var_1799_cast_fp16, y = block_cache_map)[name = string("op_1800_cast_fp16")];
            tensor<int32, [1]> nv_full_5_axes_0 = const()[name = string("nv_full_5_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_5_cast_fp16 = expand_dims(axes = nv_full_5_axes_0, x = var_1800_cast_fp16)[name = string("nv_full_5_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1806_cast_fp16 = mul(x = lkc_9_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1806_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1807_cast_fp16 = mul(x = nk_full_5_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1807_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_11_cast_fp16 = add(x = var_1806_cast_fp16, y = var_1807_cast_fp16)[name = string("lkc_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1813_cast_fp16 = mul(x = lvc_9_cast_fp16, y = var_1189_cast_fp16)[name = string("op_1813_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_1814_cast_fp16 = mul(x = nv_full_5_cast_fp16, y = update_mask_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_11_cast_fp16 = add(x = var_1813_cast_fp16, y = var_1814_cast_fp16)[name = string("lvc_11_cast_fp16")];
            tensor<int32, [1]> var_1818_axes_0 = const()[name = string("op_1818_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1818_cast_fp16 = squeeze(axes = var_1818_axes_0, x = lkc_11_cast_fp16)[name = string("op_1818_cast_fp16")];
            tensor<int32, [4]> var_1822 = const()[name = string("op_1822"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_5_cast_fp16 = reshape(shape = var_1822, x = var_1818_cast_fp16)[name = string("kc_5_cast_fp16")];
            tensor<int32, [1]> var_1825_axes_0 = const()[name = string("op_1825_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_1825_cast_fp16 = squeeze(axes = var_1825_axes_0, x = lvc_11_cast_fp16)[name = string("op_1825_cast_fp16")];
            tensor<int32, [4]> var_1829 = const()[name = string("op_1829"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_5_cast_fp16 = reshape(shape = var_1829, x = var_1825_cast_fp16)[name = string("vc_5_cast_fp16")];
            bool var_1831_transpose_x_0 = const()[name = string("op_1831_transpose_x_0"), val = bool(false)];
            bool var_1831_transpose_y_0 = const()[name = string("op_1831_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_1831_cast_fp16 = matmul(transpose_x = var_1831_transpose_x_0, transpose_y = var_1831_transpose_y_0, x = q_17_cast_fp16, y = kc_5_cast_fp16)[name = string("op_1831_cast_fp16")];
            fp16 _inversed_scores_13_y_0_to_fp16 = const()[name = string("_inversed_scores_13_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_13_cast_fp16 = mul(x = var_1831_cast_fp16, y = _inversed_scores_13_y_0_to_fp16)[name = string("_inversed_scores_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_15_cast_fp16 = add(x = _inversed_scores_13_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_15_cast_fp16")];
            int32 var_1860 = const()[name = string("op_1860"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_1862_cast_fp16 = softmax(axis = var_1860, x = scores_15_cast_fp16)[name = string("op_1862_cast_fp16")];
            bool attn_out_9_transpose_x_1 = const()[name = string("attn_out_9_transpose_x_1"), val = bool(false)];
            bool attn_out_9_transpose_y_1 = const()[name = string("attn_out_9_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_9_cast_fp16 = matmul(transpose_x = attn_out_9_transpose_x_1, transpose_y = attn_out_9_transpose_y_1, x = var_1862_cast_fp16, y = vc_5_cast_fp16)[name = string("attn_out_9_cast_fp16")];
            tensor<int32, [4]> var_1874_perm_0 = const()[name = string("op_1874_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1876 = const()[name = string("op_1876"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_1874_cast_fp16 = transpose(perm = var_1874_perm_0, x = attn_out_9_cast_fp16)[name = string("transpose_165")];
            tensor<fp16, [1, 16, 1024]> input_21_cast_fp16 = reshape(shape = var_1876, x = var_1874_cast_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_2_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_2_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(98602816)))];
            tensor<fp16, [1, 16, 1024]> linear_17_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_self_attn_o_proj_weight_to_fp16, x = input_21_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_39_cast_fp16 = add(x = hidden_states_31_cast_fp16, y = linear_17_cast_fp16)[name = string("hidden_states_39_cast_fp16")];
            fp16 var_1886_promoted_to_fp16 = const()[name = string("op_1886_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1892_cast_fp16 = pow(x = hidden_states_39_cast_fp16, y = var_1886_promoted_to_fp16)[name = string("op_1892_cast_fp16")];
            tensor<int32, [1]> variance_11_axes_0 = const()[name = string("variance_11_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_11_keep_dims_0 = const()[name = string("variance_11_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_11_cast_fp16 = reduce_mean(axes = variance_11_axes_0, keep_dims = variance_11_keep_dims_0, x = var_1892_cast_fp16)[name = string("variance_11_cast_fp16")];
            fp16 var_1895_to_fp16 = const()[name = string("op_1895_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1896_cast_fp16 = add(x = variance_11_cast_fp16, y = var_1895_to_fp16)[name = string("op_1896_cast_fp16")];
            fp32 var_1897_epsilon_0 = const()[name = string("op_1897_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1897_cast_fp16 = rsqrt(epsilon = var_1897_epsilon_0, x = var_1896_cast_fp16)[name = string("op_1897_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_43_cast_fp16 = mul(x = hidden_states_39_cast_fp16, y = var_1897_cast_fp16)[name = string("hidden_states_43_cast_fp16")];
            tensor<fp16, [1024]> layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100700032)))];
            tensor<fp16, [1, 16, 1024]> input_23_cast_fp16 = mul(x = layers_2_post_attention_layernorm_weight_to_fp16, y = hidden_states_43_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_2_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(100702144)))];
            tensor<fp16, [1, 16, 4096]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_gate_proj_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_1910_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_1910_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_2_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(109090816)))];
            tensor<fp16, [1, 16, 4096]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_2_mlp_up_proj_weight_to_fp16, x = input_23_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_27_cast_fp16 = mul(x = var_1910_cast_fp16, y = linear_19_cast_fp16)[name = string("input_27_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_2_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_2_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(117479488)))];
            tensor<fp16, [1, 16, 1024]> linear_20_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_2_mlp_down_proj_weight_to_fp16, x = input_27_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_47_cast_fp16 = add(x = hidden_states_39_cast_fp16, y = linear_20_cast_fp16)[name = string("hidden_states_47_cast_fp16")];
            tensor<int32, [4]> lkc_13_begin_0 = const()[name = string("lkc_13_begin_0"), val = tensor<int32, [4]>([0, 3072, 0, 0])];
            tensor<int32, [4]> lkc_13_end_0 = const()[name = string("lkc_13_end_0"), val = tensor<int32, [4]>([1, 4096, 1, 1024])];
            tensor<bool, [4]> lkc_13_end_mask_0 = const()[name = string("lkc_13_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_13_cast_fp16 = slice_by_index(begin = lkc_13_begin_0, end = lkc_13_end_0, end_mask = lkc_13_end_mask_0, x = key_cache)[name = string("lkc_13_cast_fp16")];
            tensor<int32, [4]> lvc_13_begin_0 = const()[name = string("lvc_13_begin_0"), val = tensor<int32, [4]>([0, 3072, 0, 0])];
            tensor<int32, [4]> lvc_13_end_0 = const()[name = string("lvc_13_end_0"), val = tensor<int32, [4]>([1, 4096, 1, 1024])];
            tensor<bool, [4]> lvc_13_end_mask_0 = const()[name = string("lvc_13_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_13_cast_fp16 = slice_by_index(begin = lvc_13_begin_0, end = lvc_13_end_0, end_mask = lvc_13_end_mask_0, x = value_cache)[name = string("lvc_13_cast_fp16")];
            fp16 var_1941_promoted_to_fp16 = const()[name = string("op_1941_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1947_cast_fp16 = pow(x = hidden_states_47_cast_fp16, y = var_1941_promoted_to_fp16)[name = string("op_1947_cast_fp16")];
            tensor<int32, [1]> variance_13_axes_0 = const()[name = string("variance_13_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_13_keep_dims_0 = const()[name = string("variance_13_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_13_cast_fp16 = reduce_mean(axes = variance_13_axes_0, keep_dims = variance_13_keep_dims_0, x = var_1947_cast_fp16)[name = string("variance_13_cast_fp16")];
            fp16 var_1950_to_fp16 = const()[name = string("op_1950_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_1951_cast_fp16 = add(x = variance_13_cast_fp16, y = var_1950_to_fp16)[name = string("op_1951_cast_fp16")];
            fp32 var_1952_epsilon_0 = const()[name = string("op_1952_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1952_cast_fp16 = rsqrt(epsilon = var_1952_epsilon_0, x = var_1951_cast_fp16)[name = string("op_1952_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_51_cast_fp16 = mul(x = hidden_states_47_cast_fp16, y = var_1952_cast_fp16)[name = string("hidden_states_51_cast_fp16")];
            tensor<fp16, [1024]> layers_3_input_layernorm_weight_to_fp16 = const()[name = string("layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125868160)))];
            tensor<fp16, [1, 16, 1024]> hidden_13_cast_fp16 = mul(x = layers_3_input_layernorm_weight_to_fp16, y = hidden_states_51_cast_fp16)[name = string("hidden_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(125870272)))];
            tensor<fp16, [1, 16, 1024]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_q_proj_weight_to_fp16, x = hidden_13_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_1976 = const()[name = string("op_1976"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_19_cast_fp16 = reshape(shape = var_1976, x = linear_21_cast_fp16)[name = string("q_19_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(127967488)))];
            tensor<fp16, [1, 16, 1024]> linear_22_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_k_proj_weight_to_fp16, x = hidden_13_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_1983 = const()[name = string("op_1983"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_13_cast_fp16 = reshape(shape = var_1983, x = linear_22_cast_fp16)[name = string("k_13_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(130064704)))];
            tensor<fp16, [1, 16, 1024]> linear_23_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_v_proj_weight_to_fp16, x = hidden_13_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_1990 = const()[name = string("op_1990"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_7_cast_fp16 = reshape(shape = var_1990, x = linear_23_cast_fp16)[name = string("v_7_cast_fp16")];
            tensor<int32, [4]> q_21_perm_0 = const()[name = string("q_21_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_15_perm_0 = const()[name = string("k_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_3_perm_0 = const()[name = string("transpose_3_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_21_cast_fp16 = transpose(perm = q_21_perm_0, x = q_19_cast_fp16)[name = string("transpose_164")];
            tensor<fp16, [1, 16, 16, 64]> var_2003_cast_fp16 = mul(x = q_21_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2003_cast_fp16")];
            tensor<int32, [4]> x1_13_begin_0 = const()[name = string("x1_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_13_end_0 = const()[name = string("x1_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_13_end_mask_0 = const()[name = string("x1_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_13_cast_fp16 = slice_by_index(begin = x1_13_begin_0, end = x1_13_end_0, end_mask = x1_13_end_mask_0, x = q_21_cast_fp16)[name = string("x1_13_cast_fp16")];
            tensor<int32, [4]> x2_13_begin_0 = const()[name = string("x2_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_13_end_0 = const()[name = string("x2_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_13_end_mask_0 = const()[name = string("x2_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_13_cast_fp16 = slice_by_index(begin = x2_13_begin_0, end = x2_13_end_0, end_mask = x2_13_end_mask_0, x = q_21_cast_fp16)[name = string("x2_13_cast_fp16")];
            fp16 const_43_promoted_to_fp16 = const()[name = string("const_43_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2024_cast_fp16 = mul(x = x2_13_cast_fp16, y = const_43_promoted_to_fp16)[name = string("op_2024_cast_fp16")];
            int32 var_2026 = const()[name = string("op_2026"), val = int32(-1)];
            bool var_2027_interleave_0 = const()[name = string("op_2027_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2027_cast_fp16 = concat(axis = var_2026, interleave = var_2027_interleave_0, values = (var_2024_cast_fp16, x1_13_cast_fp16))[name = string("op_2027_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2030_cast_fp16 = mul(x = var_2027_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2030_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_23_cast_fp16 = add(x = var_2003_cast_fp16, y = var_2030_cast_fp16)[name = string("q_23_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_15_cast_fp16 = transpose(perm = k_15_perm_0, x = k_13_cast_fp16)[name = string("transpose_163")];
            tensor<fp16, [1, 16, 16, 64]> var_2035_cast_fp16 = mul(x = k_15_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2035_cast_fp16")];
            tensor<int32, [4]> x1_15_begin_0 = const()[name = string("x1_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_15_end_0 = const()[name = string("x1_15_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_15_end_mask_0 = const()[name = string("x1_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_15_cast_fp16 = slice_by_index(begin = x1_15_begin_0, end = x1_15_end_0, end_mask = x1_15_end_mask_0, x = k_15_cast_fp16)[name = string("x1_15_cast_fp16")];
            tensor<int32, [4]> x2_15_begin_0 = const()[name = string("x2_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_15_end_0 = const()[name = string("x2_15_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_15_end_mask_0 = const()[name = string("x2_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_15_cast_fp16 = slice_by_index(begin = x2_15_begin_0, end = x2_15_end_0, end_mask = x2_15_end_mask_0, x = k_15_cast_fp16)[name = string("x2_15_cast_fp16")];
            fp16 const_46_promoted_to_fp16 = const()[name = string("const_46_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2056_cast_fp16 = mul(x = x2_15_cast_fp16, y = const_46_promoted_to_fp16)[name = string("op_2056_cast_fp16")];
            int32 var_2058 = const()[name = string("op_2058"), val = int32(-1)];
            bool var_2059_interleave_0 = const()[name = string("op_2059_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2059_cast_fp16 = concat(axis = var_2058, interleave = var_2059_interleave_0, values = (var_2056_cast_fp16, x1_15_cast_fp16))[name = string("op_2059_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2062_cast_fp16 = mul(x = var_2059_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2062_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_13_cast_fp16 = add(x = var_2035_cast_fp16, y = var_2062_cast_fp16)[name = string("x_13_cast_fp16")];
            tensor<int32, [4]> var_2069 = const()[name = string("op_2069"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2081 = const()[name = string("op_2081"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_2070_cast_fp16 = transpose(perm = var_2069, x = x_13_cast_fp16)[name = string("transpose_161")];
            tensor<fp16, [1, 1024, 1, 16]> var_2082_cast_fp16 = reshape(shape = var_2081, x = var_2070_cast_fp16)[name = string("op_2082_cast_fp16")];
            tensor<int32, [1]> var_2084_axes_0 = const()[name = string("op_2084_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2084_cast_fp16 = squeeze(axes = var_2084_axes_0, x = var_2082_cast_fp16)[name = string("op_2084_cast_fp16")];
            bool var_2085_transpose_x_0 = const()[name = string("op_2085_transpose_x_0"), val = bool(false)];
            bool var_2085_transpose_y_0 = const()[name = string("op_2085_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2085_cast_fp16 = matmul(transpose_x = var_2085_transpose_x_0, transpose_y = var_2085_transpose_y_0, x = var_2084_cast_fp16, y = block_cache_map)[name = string("op_2085_cast_fp16")];
            tensor<int32, [1]> nk_full_7_axes_0 = const()[name = string("nk_full_7_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_7_cast_fp16 = expand_dims(axes = nk_full_7_axes_0, x = var_2085_cast_fp16)[name = string("nk_full_7_cast_fp16")];
            tensor<int32, [4]> var_2104 = const()[name = string("op_2104"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_3_cast_fp16 = transpose(perm = transpose_3_perm_0, x = v_7_cast_fp16)[name = string("transpose_162")];
            tensor<fp16, [1, 1024, 1, 16]> var_2105_cast_fp16 = reshape(shape = var_2104, x = transpose_3_cast_fp16)[name = string("op_2105_cast_fp16")];
            tensor<int32, [1]> var_2107_axes_0 = const()[name = string("op_2107_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2107_cast_fp16 = squeeze(axes = var_2107_axes_0, x = var_2105_cast_fp16)[name = string("op_2107_cast_fp16")];
            bool var_2108_transpose_x_0 = const()[name = string("op_2108_transpose_x_0"), val = bool(false)];
            bool var_2108_transpose_y_0 = const()[name = string("op_2108_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2108_cast_fp16 = matmul(transpose_x = var_2108_transpose_x_0, transpose_y = var_2108_transpose_y_0, x = var_2107_cast_fp16, y = block_cache_map)[name = string("op_2108_cast_fp16")];
            tensor<int32, [1]> nv_full_7_axes_0 = const()[name = string("nv_full_7_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_7_cast_fp16 = expand_dims(axes = nv_full_7_axes_0, x = var_2108_cast_fp16)[name = string("nv_full_7_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2114_cast_fp16 = mul(x = lkc_13_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2114_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2115_cast_fp16 = mul(x = nk_full_7_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_15_cast_fp16 = add(x = var_2114_cast_fp16, y = var_2115_cast_fp16)[name = string("lkc_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2121_cast_fp16 = mul(x = lvc_13_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2121_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2122_cast_fp16 = mul(x = nv_full_7_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2122_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_15_cast_fp16 = add(x = var_2121_cast_fp16, y = var_2122_cast_fp16)[name = string("lvc_15_cast_fp16")];
            tensor<int32, [1]> var_2126_axes_0 = const()[name = string("op_2126_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2126_cast_fp16 = squeeze(axes = var_2126_axes_0, x = lkc_15_cast_fp16)[name = string("op_2126_cast_fp16")];
            tensor<int32, [4]> var_2130 = const()[name = string("op_2130"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_7_cast_fp16 = reshape(shape = var_2130, x = var_2126_cast_fp16)[name = string("kc_7_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2133_cast_fp16 = squeeze(axes = var_2133_axes_0, x = lvc_15_cast_fp16)[name = string("op_2133_cast_fp16")];
            tensor<int32, [4]> var_2137 = const()[name = string("op_2137"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_7_cast_fp16 = reshape(shape = var_2137, x = var_2133_cast_fp16)[name = string("vc_7_cast_fp16")];
            bool var_2139_transpose_x_0 = const()[name = string("op_2139_transpose_x_0"), val = bool(false)];
            bool var_2139_transpose_y_0 = const()[name = string("op_2139_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_2139_cast_fp16 = matmul(transpose_x = var_2139_transpose_x_0, transpose_y = var_2139_transpose_y_0, x = q_23_cast_fp16, y = kc_7_cast_fp16)[name = string("op_2139_cast_fp16")];
            fp16 _inversed_scores_19_y_0_to_fp16 = const()[name = string("_inversed_scores_19_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_19_cast_fp16 = mul(x = var_2139_cast_fp16, y = _inversed_scores_19_y_0_to_fp16)[name = string("_inversed_scores_19_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_21_cast_fp16 = add(x = _inversed_scores_19_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_21_cast_fp16")];
            int32 var_2168 = const()[name = string("op_2168"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_2170_cast_fp16 = softmax(axis = var_2168, x = scores_21_cast_fp16)[name = string("op_2170_cast_fp16")];
            bool attn_out_13_transpose_x_1 = const()[name = string("attn_out_13_transpose_x_1"), val = bool(false)];
            bool attn_out_13_transpose_y_1 = const()[name = string("attn_out_13_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_13_cast_fp16 = matmul(transpose_x = attn_out_13_transpose_x_1, transpose_y = attn_out_13_transpose_y_1, x = var_2170_cast_fp16, y = vc_7_cast_fp16)[name = string("attn_out_13_cast_fp16")];
            tensor<int32, [4]> var_2182_perm_0 = const()[name = string("op_2182_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2184 = const()[name = string("op_2184"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_2182_cast_fp16 = transpose(perm = var_2182_perm_0, x = attn_out_13_cast_fp16)[name = string("transpose_160")];
            tensor<fp16, [1, 16, 1024]> input_29_cast_fp16 = reshape(shape = var_2184, x = var_2182_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_3_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_3_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(132161920)))];
            tensor<fp16, [1, 16, 1024]> linear_24_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_self_attn_o_proj_weight_to_fp16, x = input_29_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_55_cast_fp16 = add(x = hidden_states_47_cast_fp16, y = linear_24_cast_fp16)[name = string("hidden_states_55_cast_fp16")];
            fp16 var_2194_promoted_to_fp16 = const()[name = string("op_2194_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2200_cast_fp16 = pow(x = hidden_states_55_cast_fp16, y = var_2194_promoted_to_fp16)[name = string("op_2200_cast_fp16")];
            tensor<int32, [1]> variance_15_axes_0 = const()[name = string("variance_15_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_15_keep_dims_0 = const()[name = string("variance_15_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_15_cast_fp16 = reduce_mean(axes = variance_15_axes_0, keep_dims = variance_15_keep_dims_0, x = var_2200_cast_fp16)[name = string("variance_15_cast_fp16")];
            fp16 var_2203_to_fp16 = const()[name = string("op_2203_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2204_cast_fp16 = add(x = variance_15_cast_fp16, y = var_2203_to_fp16)[name = string("op_2204_cast_fp16")];
            fp32 var_2205_epsilon_0 = const()[name = string("op_2205_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2205_cast_fp16 = rsqrt(epsilon = var_2205_epsilon_0, x = var_2204_cast_fp16)[name = string("op_2205_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_59_cast_fp16 = mul(x = hidden_states_55_cast_fp16, y = var_2205_cast_fp16)[name = string("hidden_states_59_cast_fp16")];
            tensor<fp16, [1024]> layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134259136)))];
            tensor<fp16, [1, 16, 1024]> input_31_cast_fp16 = mul(x = layers_3_post_attention_layernorm_weight_to_fp16, y = hidden_states_59_cast_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_3_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(134261248)))];
            tensor<fp16, [1, 16, 4096]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_gate_proj_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_2218_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_3_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(142649920)))];
            tensor<fp16, [1, 16, 4096]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_3_mlp_up_proj_weight_to_fp16, x = input_31_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_35_cast_fp16 = mul(x = var_2218_cast_fp16, y = linear_26_cast_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_3_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_3_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(151038592)))];
            tensor<fp16, [1, 16, 1024]> linear_27_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_3_mlp_down_proj_weight_to_fp16, x = input_35_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_63_cast_fp16 = add(x = hidden_states_55_cast_fp16, y = linear_27_cast_fp16)[name = string("hidden_states_63_cast_fp16")];
            tensor<int32, [4]> lkc_17_begin_0 = const()[name = string("lkc_17_begin_0"), val = tensor<int32, [4]>([0, 4096, 0, 0])];
            tensor<int32, [4]> lkc_17_end_0 = const()[name = string("lkc_17_end_0"), val = tensor<int32, [4]>([1, 5120, 1, 1024])];
            tensor<bool, [4]> lkc_17_end_mask_0 = const()[name = string("lkc_17_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_17_cast_fp16 = slice_by_index(begin = lkc_17_begin_0, end = lkc_17_end_0, end_mask = lkc_17_end_mask_0, x = key_cache)[name = string("lkc_17_cast_fp16")];
            tensor<int32, [4]> lvc_17_begin_0 = const()[name = string("lvc_17_begin_0"), val = tensor<int32, [4]>([0, 4096, 0, 0])];
            tensor<int32, [4]> lvc_17_end_0 = const()[name = string("lvc_17_end_0"), val = tensor<int32, [4]>([1, 5120, 1, 1024])];
            tensor<bool, [4]> lvc_17_end_mask_0 = const()[name = string("lvc_17_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_17_cast_fp16 = slice_by_index(begin = lvc_17_begin_0, end = lvc_17_end_0, end_mask = lvc_17_end_mask_0, x = value_cache)[name = string("lvc_17_cast_fp16")];
            fp16 var_2249_promoted_to_fp16 = const()[name = string("op_2249_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2255_cast_fp16 = pow(x = hidden_states_63_cast_fp16, y = var_2249_promoted_to_fp16)[name = string("op_2255_cast_fp16")];
            tensor<int32, [1]> variance_17_axes_0 = const()[name = string("variance_17_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_17_keep_dims_0 = const()[name = string("variance_17_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_17_cast_fp16 = reduce_mean(axes = variance_17_axes_0, keep_dims = variance_17_keep_dims_0, x = var_2255_cast_fp16)[name = string("variance_17_cast_fp16")];
            fp16 var_2258_to_fp16 = const()[name = string("op_2258_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2259_cast_fp16 = add(x = variance_17_cast_fp16, y = var_2258_to_fp16)[name = string("op_2259_cast_fp16")];
            fp32 var_2260_epsilon_0 = const()[name = string("op_2260_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2260_cast_fp16 = rsqrt(epsilon = var_2260_epsilon_0, x = var_2259_cast_fp16)[name = string("op_2260_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_67_cast_fp16 = mul(x = hidden_states_63_cast_fp16, y = var_2260_cast_fp16)[name = string("hidden_states_67_cast_fp16")];
            tensor<fp16, [1024]> layers_4_input_layernorm_weight_to_fp16 = const()[name = string("layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159427264)))];
            tensor<fp16, [1, 16, 1024]> hidden_17_cast_fp16 = mul(x = layers_4_input_layernorm_weight_to_fp16, y = hidden_states_67_cast_fp16)[name = string("hidden_17_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(159429376)))];
            tensor<fp16, [1, 16, 1024]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_q_proj_weight_to_fp16, x = hidden_17_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_2284 = const()[name = string("op_2284"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_25_cast_fp16 = reshape(shape = var_2284, x = linear_28_cast_fp16)[name = string("q_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(161526592)))];
            tensor<fp16, [1, 16, 1024]> linear_29_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_k_proj_weight_to_fp16, x = hidden_17_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2291 = const()[name = string("op_2291"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_17_cast_fp16 = reshape(shape = var_2291, x = linear_29_cast_fp16)[name = string("k_17_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(163623808)))];
            tensor<fp16, [1, 16, 1024]> linear_30_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_v_proj_weight_to_fp16, x = hidden_17_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2298 = const()[name = string("op_2298"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_9_cast_fp16 = reshape(shape = var_2298, x = linear_30_cast_fp16)[name = string("v_9_cast_fp16")];
            tensor<int32, [4]> q_27_perm_0 = const()[name = string("q_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_19_perm_0 = const()[name = string("k_19_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_4_perm_0 = const()[name = string("transpose_4_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_27_cast_fp16 = transpose(perm = q_27_perm_0, x = q_25_cast_fp16)[name = string("transpose_159")];
            tensor<fp16, [1, 16, 16, 64]> var_2311_cast_fp16 = mul(x = q_27_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2311_cast_fp16")];
            tensor<int32, [4]> x1_17_begin_0 = const()[name = string("x1_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_17_end_0 = const()[name = string("x1_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_17_end_mask_0 = const()[name = string("x1_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_17_cast_fp16 = slice_by_index(begin = x1_17_begin_0, end = x1_17_end_0, end_mask = x1_17_end_mask_0, x = q_27_cast_fp16)[name = string("x1_17_cast_fp16")];
            tensor<int32, [4]> x2_17_begin_0 = const()[name = string("x2_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_17_end_0 = const()[name = string("x2_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_17_end_mask_0 = const()[name = string("x2_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_17_cast_fp16 = slice_by_index(begin = x2_17_begin_0, end = x2_17_end_0, end_mask = x2_17_end_mask_0, x = q_27_cast_fp16)[name = string("x2_17_cast_fp16")];
            fp16 const_55_promoted_to_fp16 = const()[name = string("const_55_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2332_cast_fp16 = mul(x = x2_17_cast_fp16, y = const_55_promoted_to_fp16)[name = string("op_2332_cast_fp16")];
            int32 var_2334 = const()[name = string("op_2334"), val = int32(-1)];
            bool var_2335_interleave_0 = const()[name = string("op_2335_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2335_cast_fp16 = concat(axis = var_2334, interleave = var_2335_interleave_0, values = (var_2332_cast_fp16, x1_17_cast_fp16))[name = string("op_2335_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2338_cast_fp16 = mul(x = var_2335_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_29_cast_fp16 = add(x = var_2311_cast_fp16, y = var_2338_cast_fp16)[name = string("q_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_19_cast_fp16 = transpose(perm = k_19_perm_0, x = k_17_cast_fp16)[name = string("transpose_158")];
            tensor<fp16, [1, 16, 16, 64]> var_2343_cast_fp16 = mul(x = k_19_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2343_cast_fp16")];
            tensor<int32, [4]> x1_19_begin_0 = const()[name = string("x1_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_19_end_0 = const()[name = string("x1_19_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_19_end_mask_0 = const()[name = string("x1_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_19_cast_fp16 = slice_by_index(begin = x1_19_begin_0, end = x1_19_end_0, end_mask = x1_19_end_mask_0, x = k_19_cast_fp16)[name = string("x1_19_cast_fp16")];
            tensor<int32, [4]> x2_19_begin_0 = const()[name = string("x2_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_19_end_0 = const()[name = string("x2_19_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_19_end_mask_0 = const()[name = string("x2_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_19_cast_fp16 = slice_by_index(begin = x2_19_begin_0, end = x2_19_end_0, end_mask = x2_19_end_mask_0, x = k_19_cast_fp16)[name = string("x2_19_cast_fp16")];
            fp16 const_58_promoted_to_fp16 = const()[name = string("const_58_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2364_cast_fp16 = mul(x = x2_19_cast_fp16, y = const_58_promoted_to_fp16)[name = string("op_2364_cast_fp16")];
            int32 var_2366 = const()[name = string("op_2366"), val = int32(-1)];
            bool var_2367_interleave_0 = const()[name = string("op_2367_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2367_cast_fp16 = concat(axis = var_2366, interleave = var_2367_interleave_0, values = (var_2364_cast_fp16, x1_19_cast_fp16))[name = string("op_2367_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2370_cast_fp16 = mul(x = var_2367_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2370_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_17_cast_fp16 = add(x = var_2343_cast_fp16, y = var_2370_cast_fp16)[name = string("x_17_cast_fp16")];
            tensor<int32, [4]> var_2377 = const()[name = string("op_2377"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2389 = const()[name = string("op_2389"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_2378_cast_fp16 = transpose(perm = var_2377, x = x_17_cast_fp16)[name = string("transpose_156")];
            tensor<fp16, [1, 1024, 1, 16]> var_2390_cast_fp16 = reshape(shape = var_2389, x = var_2378_cast_fp16)[name = string("op_2390_cast_fp16")];
            tensor<int32, [1]> var_2392_axes_0 = const()[name = string("op_2392_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2392_cast_fp16 = squeeze(axes = var_2392_axes_0, x = var_2390_cast_fp16)[name = string("op_2392_cast_fp16")];
            bool var_2393_transpose_x_0 = const()[name = string("op_2393_transpose_x_0"), val = bool(false)];
            bool var_2393_transpose_y_0 = const()[name = string("op_2393_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2393_cast_fp16 = matmul(transpose_x = var_2393_transpose_x_0, transpose_y = var_2393_transpose_y_0, x = var_2392_cast_fp16, y = block_cache_map)[name = string("op_2393_cast_fp16")];
            tensor<int32, [1]> nk_full_9_axes_0 = const()[name = string("nk_full_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_9_cast_fp16 = expand_dims(axes = nk_full_9_axes_0, x = var_2393_cast_fp16)[name = string("nk_full_9_cast_fp16")];
            tensor<int32, [4]> var_2412 = const()[name = string("op_2412"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_4_cast_fp16 = transpose(perm = transpose_4_perm_0, x = v_9_cast_fp16)[name = string("transpose_157")];
            tensor<fp16, [1, 1024, 1, 16]> var_2413_cast_fp16 = reshape(shape = var_2412, x = transpose_4_cast_fp16)[name = string("op_2413_cast_fp16")];
            tensor<int32, [1]> var_2415_axes_0 = const()[name = string("op_2415_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2415_cast_fp16 = squeeze(axes = var_2415_axes_0, x = var_2413_cast_fp16)[name = string("op_2415_cast_fp16")];
            bool var_2416_transpose_x_0 = const()[name = string("op_2416_transpose_x_0"), val = bool(false)];
            bool var_2416_transpose_y_0 = const()[name = string("op_2416_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2416_cast_fp16 = matmul(transpose_x = var_2416_transpose_x_0, transpose_y = var_2416_transpose_y_0, x = var_2415_cast_fp16, y = block_cache_map)[name = string("op_2416_cast_fp16")];
            tensor<int32, [1]> nv_full_9_axes_0 = const()[name = string("nv_full_9_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_9_cast_fp16 = expand_dims(axes = nv_full_9_axes_0, x = var_2416_cast_fp16)[name = string("nv_full_9_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2422_cast_fp16 = mul(x = lkc_17_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2422_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2423_cast_fp16 = mul(x = nk_full_9_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2423_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_19_cast_fp16 = add(x = var_2422_cast_fp16, y = var_2423_cast_fp16)[name = string("lkc_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2429_cast_fp16 = mul(x = lvc_17_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2429_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2430_cast_fp16 = mul(x = nv_full_9_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2430_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_19_cast_fp16 = add(x = var_2429_cast_fp16, y = var_2430_cast_fp16)[name = string("lvc_19_cast_fp16")];
            tensor<int32, [1]> var_2434_axes_0 = const()[name = string("op_2434_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2434_cast_fp16 = squeeze(axes = var_2434_axes_0, x = lkc_19_cast_fp16)[name = string("op_2434_cast_fp16")];
            tensor<int32, [4]> var_2438 = const()[name = string("op_2438"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_9_cast_fp16 = reshape(shape = var_2438, x = var_2434_cast_fp16)[name = string("kc_9_cast_fp16")];
            tensor<int32, [1]> var_2441_axes_0 = const()[name = string("op_2441_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2441_cast_fp16 = squeeze(axes = var_2441_axes_0, x = lvc_19_cast_fp16)[name = string("op_2441_cast_fp16")];
            tensor<int32, [4]> var_2445 = const()[name = string("op_2445"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_9_cast_fp16 = reshape(shape = var_2445, x = var_2441_cast_fp16)[name = string("vc_9_cast_fp16")];
            bool var_2447_transpose_x_0 = const()[name = string("op_2447_transpose_x_0"), val = bool(false)];
            bool var_2447_transpose_y_0 = const()[name = string("op_2447_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_2447_cast_fp16 = matmul(transpose_x = var_2447_transpose_x_0, transpose_y = var_2447_transpose_y_0, x = q_29_cast_fp16, y = kc_9_cast_fp16)[name = string("op_2447_cast_fp16")];
            fp16 _inversed_scores_25_y_0_to_fp16 = const()[name = string("_inversed_scores_25_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_25_cast_fp16 = mul(x = var_2447_cast_fp16, y = _inversed_scores_25_y_0_to_fp16)[name = string("_inversed_scores_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_27_cast_fp16 = add(x = _inversed_scores_25_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_27_cast_fp16")];
            int32 var_2476 = const()[name = string("op_2476"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_2478_cast_fp16 = softmax(axis = var_2476, x = scores_27_cast_fp16)[name = string("op_2478_cast_fp16")];
            bool attn_out_17_transpose_x_1 = const()[name = string("attn_out_17_transpose_x_1"), val = bool(false)];
            bool attn_out_17_transpose_y_1 = const()[name = string("attn_out_17_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_17_cast_fp16 = matmul(transpose_x = attn_out_17_transpose_x_1, transpose_y = attn_out_17_transpose_y_1, x = var_2478_cast_fp16, y = vc_9_cast_fp16)[name = string("attn_out_17_cast_fp16")];
            tensor<int32, [4]> var_2490_perm_0 = const()[name = string("op_2490_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2492 = const()[name = string("op_2492"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_2490_cast_fp16 = transpose(perm = var_2490_perm_0, x = attn_out_17_cast_fp16)[name = string("transpose_155")];
            tensor<fp16, [1, 16, 1024]> input_37_cast_fp16 = reshape(shape = var_2492, x = var_2490_cast_fp16)[name = string("input_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_4_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_4_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(165721024)))];
            tensor<fp16, [1, 16, 1024]> linear_31_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_self_attn_o_proj_weight_to_fp16, x = input_37_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_71_cast_fp16 = add(x = hidden_states_63_cast_fp16, y = linear_31_cast_fp16)[name = string("hidden_states_71_cast_fp16")];
            fp16 var_2502_promoted_to_fp16 = const()[name = string("op_2502_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2508_cast_fp16 = pow(x = hidden_states_71_cast_fp16, y = var_2502_promoted_to_fp16)[name = string("op_2508_cast_fp16")];
            tensor<int32, [1]> variance_19_axes_0 = const()[name = string("variance_19_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_19_keep_dims_0 = const()[name = string("variance_19_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_19_cast_fp16 = reduce_mean(axes = variance_19_axes_0, keep_dims = variance_19_keep_dims_0, x = var_2508_cast_fp16)[name = string("variance_19_cast_fp16")];
            fp16 var_2511_to_fp16 = const()[name = string("op_2511_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2512_cast_fp16 = add(x = variance_19_cast_fp16, y = var_2511_to_fp16)[name = string("op_2512_cast_fp16")];
            fp32 var_2513_epsilon_0 = const()[name = string("op_2513_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2513_cast_fp16 = rsqrt(epsilon = var_2513_epsilon_0, x = var_2512_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_75_cast_fp16 = mul(x = hidden_states_71_cast_fp16, y = var_2513_cast_fp16)[name = string("hidden_states_75_cast_fp16")];
            tensor<fp16, [1024]> layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167818240)))];
            tensor<fp16, [1, 16, 1024]> input_39_cast_fp16 = mul(x = layers_4_post_attention_layernorm_weight_to_fp16, y = hidden_states_75_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_4_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_4_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(167820352)))];
            tensor<fp16, [1, 16, 4096]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_gate_proj_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_2526_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_2526_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_4_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_4_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(176209024)))];
            tensor<fp16, [1, 16, 4096]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_4_mlp_up_proj_weight_to_fp16, x = input_39_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_43_cast_fp16 = mul(x = var_2526_cast_fp16, y = linear_33_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_4_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_4_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(184597696)))];
            tensor<fp16, [1, 16, 1024]> linear_34_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_4_mlp_down_proj_weight_to_fp16, x = input_43_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_79_cast_fp16 = add(x = hidden_states_71_cast_fp16, y = linear_34_cast_fp16)[name = string("hidden_states_79_cast_fp16")];
            tensor<int32, [4]> lkc_21_begin_0 = const()[name = string("lkc_21_begin_0"), val = tensor<int32, [4]>([0, 5120, 0, 0])];
            tensor<int32, [4]> lkc_21_end_0 = const()[name = string("lkc_21_end_0"), val = tensor<int32, [4]>([1, 6144, 1, 1024])];
            tensor<bool, [4]> lkc_21_end_mask_0 = const()[name = string("lkc_21_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_21_cast_fp16 = slice_by_index(begin = lkc_21_begin_0, end = lkc_21_end_0, end_mask = lkc_21_end_mask_0, x = key_cache)[name = string("lkc_21_cast_fp16")];
            tensor<int32, [4]> lvc_21_begin_0 = const()[name = string("lvc_21_begin_0"), val = tensor<int32, [4]>([0, 5120, 0, 0])];
            tensor<int32, [4]> lvc_21_end_0 = const()[name = string("lvc_21_end_0"), val = tensor<int32, [4]>([1, 6144, 1, 1024])];
            tensor<bool, [4]> lvc_21_end_mask_0 = const()[name = string("lvc_21_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_21_cast_fp16 = slice_by_index(begin = lvc_21_begin_0, end = lvc_21_end_0, end_mask = lvc_21_end_mask_0, x = value_cache)[name = string("lvc_21_cast_fp16")];
            fp16 var_2557_promoted_to_fp16 = const()[name = string("op_2557_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2563_cast_fp16 = pow(x = hidden_states_79_cast_fp16, y = var_2557_promoted_to_fp16)[name = string("op_2563_cast_fp16")];
            tensor<int32, [1]> variance_21_axes_0 = const()[name = string("variance_21_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_21_keep_dims_0 = const()[name = string("variance_21_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_21_cast_fp16 = reduce_mean(axes = variance_21_axes_0, keep_dims = variance_21_keep_dims_0, x = var_2563_cast_fp16)[name = string("variance_21_cast_fp16")];
            fp16 var_2566_to_fp16 = const()[name = string("op_2566_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2567_cast_fp16 = add(x = variance_21_cast_fp16, y = var_2566_to_fp16)[name = string("op_2567_cast_fp16")];
            fp32 var_2568_epsilon_0 = const()[name = string("op_2568_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2568_cast_fp16 = rsqrt(epsilon = var_2568_epsilon_0, x = var_2567_cast_fp16)[name = string("op_2568_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_83_cast_fp16 = mul(x = hidden_states_79_cast_fp16, y = var_2568_cast_fp16)[name = string("hidden_states_83_cast_fp16")];
            tensor<fp16, [1024]> layers_5_input_layernorm_weight_to_fp16 = const()[name = string("layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192986368)))];
            tensor<fp16, [1, 16, 1024]> hidden_21_cast_fp16 = mul(x = layers_5_input_layernorm_weight_to_fp16, y = hidden_states_83_cast_fp16)[name = string("hidden_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(192988480)))];
            tensor<fp16, [1, 16, 1024]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_q_proj_weight_to_fp16, x = hidden_21_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_2592 = const()[name = string("op_2592"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_31_cast_fp16 = reshape(shape = var_2592, x = linear_35_cast_fp16)[name = string("q_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(195085696)))];
            tensor<fp16, [1, 16, 1024]> linear_36_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_k_proj_weight_to_fp16, x = hidden_21_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_2599 = const()[name = string("op_2599"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_21_cast_fp16 = reshape(shape = var_2599, x = linear_36_cast_fp16)[name = string("k_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(197182912)))];
            tensor<fp16, [1, 16, 1024]> linear_37_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_v_proj_weight_to_fp16, x = hidden_21_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_2606 = const()[name = string("op_2606"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_11_cast_fp16 = reshape(shape = var_2606, x = linear_37_cast_fp16)[name = string("v_11_cast_fp16")];
            tensor<int32, [4]> q_33_perm_0 = const()[name = string("q_33_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_23_perm_0 = const()[name = string("k_23_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_5_perm_0 = const()[name = string("transpose_5_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_33_cast_fp16 = transpose(perm = q_33_perm_0, x = q_31_cast_fp16)[name = string("transpose_154")];
            tensor<fp16, [1, 16, 16, 64]> var_2619_cast_fp16 = mul(x = q_33_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2619_cast_fp16")];
            tensor<int32, [4]> x1_21_begin_0 = const()[name = string("x1_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_21_end_0 = const()[name = string("x1_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_21_end_mask_0 = const()[name = string("x1_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_21_cast_fp16 = slice_by_index(begin = x1_21_begin_0, end = x1_21_end_0, end_mask = x1_21_end_mask_0, x = q_33_cast_fp16)[name = string("x1_21_cast_fp16")];
            tensor<int32, [4]> x2_21_begin_0 = const()[name = string("x2_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_21_end_0 = const()[name = string("x2_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_21_end_mask_0 = const()[name = string("x2_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_21_cast_fp16 = slice_by_index(begin = x2_21_begin_0, end = x2_21_end_0, end_mask = x2_21_end_mask_0, x = q_33_cast_fp16)[name = string("x2_21_cast_fp16")];
            fp16 const_67_promoted_to_fp16 = const()[name = string("const_67_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2640_cast_fp16 = mul(x = x2_21_cast_fp16, y = const_67_promoted_to_fp16)[name = string("op_2640_cast_fp16")];
            int32 var_2642 = const()[name = string("op_2642"), val = int32(-1)];
            bool var_2643_interleave_0 = const()[name = string("op_2643_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2643_cast_fp16 = concat(axis = var_2642, interleave = var_2643_interleave_0, values = (var_2640_cast_fp16, x1_21_cast_fp16))[name = string("op_2643_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2646_cast_fp16 = mul(x = var_2643_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2646_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_35_cast_fp16 = add(x = var_2619_cast_fp16, y = var_2646_cast_fp16)[name = string("q_35_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_23_cast_fp16 = transpose(perm = k_23_perm_0, x = k_21_cast_fp16)[name = string("transpose_153")];
            tensor<fp16, [1, 16, 16, 64]> var_2651_cast_fp16 = mul(x = k_23_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2651_cast_fp16")];
            tensor<int32, [4]> x1_23_begin_0 = const()[name = string("x1_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_23_end_0 = const()[name = string("x1_23_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_23_end_mask_0 = const()[name = string("x1_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_23_cast_fp16 = slice_by_index(begin = x1_23_begin_0, end = x1_23_end_0, end_mask = x1_23_end_mask_0, x = k_23_cast_fp16)[name = string("x1_23_cast_fp16")];
            tensor<int32, [4]> x2_23_begin_0 = const()[name = string("x2_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_23_end_0 = const()[name = string("x2_23_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_23_end_mask_0 = const()[name = string("x2_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_23_cast_fp16 = slice_by_index(begin = x2_23_begin_0, end = x2_23_end_0, end_mask = x2_23_end_mask_0, x = k_23_cast_fp16)[name = string("x2_23_cast_fp16")];
            fp16 const_70_promoted_to_fp16 = const()[name = string("const_70_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2672_cast_fp16 = mul(x = x2_23_cast_fp16, y = const_70_promoted_to_fp16)[name = string("op_2672_cast_fp16")];
            int32 var_2674 = const()[name = string("op_2674"), val = int32(-1)];
            bool var_2675_interleave_0 = const()[name = string("op_2675_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2675_cast_fp16 = concat(axis = var_2674, interleave = var_2675_interleave_0, values = (var_2672_cast_fp16, x1_23_cast_fp16))[name = string("op_2675_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2678_cast_fp16 = mul(x = var_2675_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2678_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_21_cast_fp16 = add(x = var_2651_cast_fp16, y = var_2678_cast_fp16)[name = string("x_21_cast_fp16")];
            tensor<int32, [4]> var_2685 = const()[name = string("op_2685"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_2697 = const()[name = string("op_2697"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_2686_cast_fp16 = transpose(perm = var_2685, x = x_21_cast_fp16)[name = string("transpose_151")];
            tensor<fp16, [1, 1024, 1, 16]> var_2698_cast_fp16 = reshape(shape = var_2697, x = var_2686_cast_fp16)[name = string("op_2698_cast_fp16")];
            tensor<int32, [1]> var_2700_axes_0 = const()[name = string("op_2700_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2700_cast_fp16 = squeeze(axes = var_2700_axes_0, x = var_2698_cast_fp16)[name = string("op_2700_cast_fp16")];
            bool var_2701_transpose_x_0 = const()[name = string("op_2701_transpose_x_0"), val = bool(false)];
            bool var_2701_transpose_y_0 = const()[name = string("op_2701_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2701_cast_fp16 = matmul(transpose_x = var_2701_transpose_x_0, transpose_y = var_2701_transpose_y_0, x = var_2700_cast_fp16, y = block_cache_map)[name = string("op_2701_cast_fp16")];
            tensor<int32, [1]> nk_full_11_axes_0 = const()[name = string("nk_full_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_11_cast_fp16 = expand_dims(axes = nk_full_11_axes_0, x = var_2701_cast_fp16)[name = string("nk_full_11_cast_fp16")];
            tensor<int32, [4]> var_2720 = const()[name = string("op_2720"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_5_cast_fp16 = transpose(perm = transpose_5_perm_0, x = v_11_cast_fp16)[name = string("transpose_152")];
            tensor<fp16, [1, 1024, 1, 16]> var_2721_cast_fp16 = reshape(shape = var_2720, x = transpose_5_cast_fp16)[name = string("op_2721_cast_fp16")];
            tensor<int32, [1]> var_2723_axes_0 = const()[name = string("op_2723_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_2723_cast_fp16 = squeeze(axes = var_2723_axes_0, x = var_2721_cast_fp16)[name = string("op_2723_cast_fp16")];
            bool var_2724_transpose_x_0 = const()[name = string("op_2724_transpose_x_0"), val = bool(false)];
            bool var_2724_transpose_y_0 = const()[name = string("op_2724_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_2724_cast_fp16 = matmul(transpose_x = var_2724_transpose_x_0, transpose_y = var_2724_transpose_y_0, x = var_2723_cast_fp16, y = block_cache_map)[name = string("op_2724_cast_fp16")];
            tensor<int32, [1]> nv_full_11_axes_0 = const()[name = string("nv_full_11_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_11_cast_fp16 = expand_dims(axes = nv_full_11_axes_0, x = var_2724_cast_fp16)[name = string("nv_full_11_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2730_cast_fp16 = mul(x = lkc_21_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2730_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2731_cast_fp16 = mul(x = nk_full_11_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2731_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_23_cast_fp16 = add(x = var_2730_cast_fp16, y = var_2731_cast_fp16)[name = string("lkc_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2737_cast_fp16 = mul(x = lvc_21_cast_fp16, y = var_1189_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_2738_cast_fp16 = mul(x = nv_full_11_cast_fp16, y = update_mask_cast_fp16)[name = string("op_2738_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_23_cast_fp16 = add(x = var_2737_cast_fp16, y = var_2738_cast_fp16)[name = string("lvc_23_cast_fp16")];
            tensor<int32, [1]> var_2742_axes_0 = const()[name = string("op_2742_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2742_cast_fp16 = squeeze(axes = var_2742_axes_0, x = lkc_23_cast_fp16)[name = string("op_2742_cast_fp16")];
            tensor<int32, [4]> var_2746 = const()[name = string("op_2746"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_11_cast_fp16 = reshape(shape = var_2746, x = var_2742_cast_fp16)[name = string("kc_11_cast_fp16")];
            tensor<int32, [1]> var_2749_axes_0 = const()[name = string("op_2749_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_2749_cast_fp16 = squeeze(axes = var_2749_axes_0, x = lvc_23_cast_fp16)[name = string("op_2749_cast_fp16")];
            tensor<int32, [4]> var_2753 = const()[name = string("op_2753"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_11_cast_fp16 = reshape(shape = var_2753, x = var_2749_cast_fp16)[name = string("vc_11_cast_fp16")];
            bool var_2755_transpose_x_0 = const()[name = string("op_2755_transpose_x_0"), val = bool(false)];
            bool var_2755_transpose_y_0 = const()[name = string("op_2755_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_2755_cast_fp16 = matmul(transpose_x = var_2755_transpose_x_0, transpose_y = var_2755_transpose_y_0, x = q_35_cast_fp16, y = kc_11_cast_fp16)[name = string("op_2755_cast_fp16")];
            fp16 _inversed_scores_31_y_0_to_fp16 = const()[name = string("_inversed_scores_31_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_31_cast_fp16 = mul(x = var_2755_cast_fp16, y = _inversed_scores_31_y_0_to_fp16)[name = string("_inversed_scores_31_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_33_cast_fp16 = add(x = _inversed_scores_31_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_33_cast_fp16")];
            int32 var_2784 = const()[name = string("op_2784"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_2786_cast_fp16 = softmax(axis = var_2784, x = scores_33_cast_fp16)[name = string("op_2786_cast_fp16")];
            bool attn_out_21_transpose_x_1 = const()[name = string("attn_out_21_transpose_x_1"), val = bool(false)];
            bool attn_out_21_transpose_y_1 = const()[name = string("attn_out_21_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_21_cast_fp16 = matmul(transpose_x = attn_out_21_transpose_x_1, transpose_y = attn_out_21_transpose_y_1, x = var_2786_cast_fp16, y = vc_11_cast_fp16)[name = string("attn_out_21_cast_fp16")];
            tensor<int32, [4]> var_2798_perm_0 = const()[name = string("op_2798_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2800 = const()[name = string("op_2800"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_2798_cast_fp16 = transpose(perm = var_2798_perm_0, x = attn_out_21_cast_fp16)[name = string("transpose_150")];
            tensor<fp16, [1, 16, 1024]> input_45_cast_fp16 = reshape(shape = var_2800, x = var_2798_cast_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_5_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_5_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(199280128)))];
            tensor<fp16, [1, 16, 1024]> linear_38_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_self_attn_o_proj_weight_to_fp16, x = input_45_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_87_cast_fp16 = add(x = hidden_states_79_cast_fp16, y = linear_38_cast_fp16)[name = string("hidden_states_87_cast_fp16")];
            fp16 var_2810_promoted_to_fp16 = const()[name = string("op_2810_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2816_cast_fp16 = pow(x = hidden_states_87_cast_fp16, y = var_2810_promoted_to_fp16)[name = string("op_2816_cast_fp16")];
            tensor<int32, [1]> variance_23_axes_0 = const()[name = string("variance_23_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_23_keep_dims_0 = const()[name = string("variance_23_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_23_cast_fp16 = reduce_mean(axes = variance_23_axes_0, keep_dims = variance_23_keep_dims_0, x = var_2816_cast_fp16)[name = string("variance_23_cast_fp16")];
            fp16 var_2819_to_fp16 = const()[name = string("op_2819_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2820_cast_fp16 = add(x = variance_23_cast_fp16, y = var_2819_to_fp16)[name = string("op_2820_cast_fp16")];
            fp32 var_2821_epsilon_0 = const()[name = string("op_2821_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2821_cast_fp16 = rsqrt(epsilon = var_2821_epsilon_0, x = var_2820_cast_fp16)[name = string("op_2821_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_91_cast_fp16 = mul(x = hidden_states_87_cast_fp16, y = var_2821_cast_fp16)[name = string("hidden_states_91_cast_fp16")];
            tensor<fp16, [1024]> layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201377344)))];
            tensor<fp16, [1, 16, 1024]> input_47_cast_fp16 = mul(x = layers_5_post_attention_layernorm_weight_to_fp16, y = hidden_states_91_cast_fp16)[name = string("input_47_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_5_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(201379456)))];
            tensor<fp16, [1, 16, 4096]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_gate_proj_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_2834_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_2834_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_5_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(209768128)))];
            tensor<fp16, [1, 16, 4096]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_5_mlp_up_proj_weight_to_fp16, x = input_47_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_51_cast_fp16 = mul(x = var_2834_cast_fp16, y = linear_40_cast_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_5_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_5_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(218156800)))];
            tensor<fp16, [1, 16, 1024]> linear_41_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_5_mlp_down_proj_weight_to_fp16, x = input_51_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_95_cast_fp16 = add(x = hidden_states_87_cast_fp16, y = linear_41_cast_fp16)[name = string("hidden_states_95_cast_fp16")];
            tensor<int32, [4]> lkc_25_begin_0 = const()[name = string("lkc_25_begin_0"), val = tensor<int32, [4]>([0, 6144, 0, 0])];
            tensor<int32, [4]> lkc_25_end_0 = const()[name = string("lkc_25_end_0"), val = tensor<int32, [4]>([1, 7168, 1, 1024])];
            tensor<bool, [4]> lkc_25_end_mask_0 = const()[name = string("lkc_25_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_25_cast_fp16 = slice_by_index(begin = lkc_25_begin_0, end = lkc_25_end_0, end_mask = lkc_25_end_mask_0, x = key_cache)[name = string("lkc_25_cast_fp16")];
            tensor<int32, [4]> lvc_25_begin_0 = const()[name = string("lvc_25_begin_0"), val = tensor<int32, [4]>([0, 6144, 0, 0])];
            tensor<int32, [4]> lvc_25_end_0 = const()[name = string("lvc_25_end_0"), val = tensor<int32, [4]>([1, 7168, 1, 1024])];
            tensor<bool, [4]> lvc_25_end_mask_0 = const()[name = string("lvc_25_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_25_cast_fp16 = slice_by_index(begin = lvc_25_begin_0, end = lvc_25_end_0, end_mask = lvc_25_end_mask_0, x = value_cache)[name = string("lvc_25_cast_fp16")];
            fp16 var_2865_promoted_to_fp16 = const()[name = string("op_2865_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2871_cast_fp16 = pow(x = hidden_states_95_cast_fp16, y = var_2865_promoted_to_fp16)[name = string("op_2871_cast_fp16")];
            tensor<int32, [1]> variance_25_axes_0 = const()[name = string("variance_25_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_25_keep_dims_0 = const()[name = string("variance_25_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_25_cast_fp16 = reduce_mean(axes = variance_25_axes_0, keep_dims = variance_25_keep_dims_0, x = var_2871_cast_fp16)[name = string("variance_25_cast_fp16")];
            fp16 var_2874_to_fp16 = const()[name = string("op_2874_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_2875_cast_fp16 = add(x = variance_25_cast_fp16, y = var_2874_to_fp16)[name = string("op_2875_cast_fp16")];
            fp32 var_2876_epsilon_0 = const()[name = string("op_2876_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2876_cast_fp16 = rsqrt(epsilon = var_2876_epsilon_0, x = var_2875_cast_fp16)[name = string("op_2876_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_99_cast_fp16 = mul(x = hidden_states_95_cast_fp16, y = var_2876_cast_fp16)[name = string("hidden_states_99_cast_fp16")];
            tensor<fp16, [1024]> layers_6_input_layernorm_weight_to_fp16 = const()[name = string("layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226545472)))];
            tensor<fp16, [1, 16, 1024]> hidden_25_cast_fp16 = mul(x = layers_6_input_layernorm_weight_to_fp16, y = hidden_states_99_cast_fp16)[name = string("hidden_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(226547584)))];
            tensor<fp16, [1, 16, 1024]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_q_proj_weight_to_fp16, x = hidden_25_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_2900 = const()[name = string("op_2900"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_37_cast_fp16 = reshape(shape = var_2900, x = linear_42_cast_fp16)[name = string("q_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(228644800)))];
            tensor<fp16, [1, 16, 1024]> linear_43_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_k_proj_weight_to_fp16, x = hidden_25_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_2907 = const()[name = string("op_2907"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_25_cast_fp16 = reshape(shape = var_2907, x = linear_43_cast_fp16)[name = string("k_25_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(230742016)))];
            tensor<fp16, [1, 16, 1024]> linear_44_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_v_proj_weight_to_fp16, x = hidden_25_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_2914 = const()[name = string("op_2914"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_13_cast_fp16 = reshape(shape = var_2914, x = linear_44_cast_fp16)[name = string("v_13_cast_fp16")];
            tensor<int32, [4]> q_39_perm_0 = const()[name = string("q_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_27_perm_0 = const()[name = string("k_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_6_perm_0 = const()[name = string("transpose_6_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_39_cast_fp16 = transpose(perm = q_39_perm_0, x = q_37_cast_fp16)[name = string("transpose_149")];
            tensor<fp16, [1, 16, 16, 64]> var_2927_cast_fp16 = mul(x = q_39_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2927_cast_fp16")];
            tensor<int32, [4]> x1_25_begin_0 = const()[name = string("x1_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_25_end_0 = const()[name = string("x1_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_25_end_mask_0 = const()[name = string("x1_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_25_cast_fp16 = slice_by_index(begin = x1_25_begin_0, end = x1_25_end_0, end_mask = x1_25_end_mask_0, x = q_39_cast_fp16)[name = string("x1_25_cast_fp16")];
            tensor<int32, [4]> x2_25_begin_0 = const()[name = string("x2_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_25_end_0 = const()[name = string("x2_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_25_end_mask_0 = const()[name = string("x2_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_25_cast_fp16 = slice_by_index(begin = x2_25_begin_0, end = x2_25_end_0, end_mask = x2_25_end_mask_0, x = q_39_cast_fp16)[name = string("x2_25_cast_fp16")];
            fp16 const_79_promoted_to_fp16 = const()[name = string("const_79_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2948_cast_fp16 = mul(x = x2_25_cast_fp16, y = const_79_promoted_to_fp16)[name = string("op_2948_cast_fp16")];
            int32 var_2950 = const()[name = string("op_2950"), val = int32(-1)];
            bool var_2951_interleave_0 = const()[name = string("op_2951_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2951_cast_fp16 = concat(axis = var_2950, interleave = var_2951_interleave_0, values = (var_2948_cast_fp16, x1_25_cast_fp16))[name = string("op_2951_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2954_cast_fp16 = mul(x = var_2951_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2954_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_41_cast_fp16 = add(x = var_2927_cast_fp16, y = var_2954_cast_fp16)[name = string("q_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_27_cast_fp16 = transpose(perm = k_27_perm_0, x = k_25_cast_fp16)[name = string("transpose_148")];
            tensor<fp16, [1, 16, 16, 64]> var_2959_cast_fp16 = mul(x = k_27_cast_fp16, y = var_1078_cast_fp16)[name = string("op_2959_cast_fp16")];
            tensor<int32, [4]> x1_27_begin_0 = const()[name = string("x1_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_27_end_0 = const()[name = string("x1_27_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_27_end_mask_0 = const()[name = string("x1_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_27_cast_fp16 = slice_by_index(begin = x1_27_begin_0, end = x1_27_end_0, end_mask = x1_27_end_mask_0, x = k_27_cast_fp16)[name = string("x1_27_cast_fp16")];
            tensor<int32, [4]> x2_27_begin_0 = const()[name = string("x2_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_27_end_0 = const()[name = string("x2_27_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_27_end_mask_0 = const()[name = string("x2_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_27_cast_fp16 = slice_by_index(begin = x2_27_begin_0, end = x2_27_end_0, end_mask = x2_27_end_mask_0, x = k_27_cast_fp16)[name = string("x2_27_cast_fp16")];
            fp16 const_82_promoted_to_fp16 = const()[name = string("const_82_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_2980_cast_fp16 = mul(x = x2_27_cast_fp16, y = const_82_promoted_to_fp16)[name = string("op_2980_cast_fp16")];
            int32 var_2982 = const()[name = string("op_2982"), val = int32(-1)];
            bool var_2983_interleave_0 = const()[name = string("op_2983_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_2983_cast_fp16 = concat(axis = var_2982, interleave = var_2983_interleave_0, values = (var_2980_cast_fp16, x1_27_cast_fp16))[name = string("op_2983_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2986_cast_fp16 = mul(x = var_2983_cast_fp16, y = var_1105_cast_fp16)[name = string("op_2986_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_25_cast_fp16 = add(x = var_2959_cast_fp16, y = var_2986_cast_fp16)[name = string("x_25_cast_fp16")];
            tensor<int32, [4]> var_2993 = const()[name = string("op_2993"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3005 = const()[name = string("op_3005"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_2994_cast_fp16 = transpose(perm = var_2993, x = x_25_cast_fp16)[name = string("transpose_146")];
            tensor<fp16, [1, 1024, 1, 16]> var_3006_cast_fp16 = reshape(shape = var_3005, x = var_2994_cast_fp16)[name = string("op_3006_cast_fp16")];
            tensor<int32, [1]> var_3008_axes_0 = const()[name = string("op_3008_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3008_cast_fp16 = squeeze(axes = var_3008_axes_0, x = var_3006_cast_fp16)[name = string("op_3008_cast_fp16")];
            bool var_3009_transpose_x_0 = const()[name = string("op_3009_transpose_x_0"), val = bool(false)];
            bool var_3009_transpose_y_0 = const()[name = string("op_3009_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3009_cast_fp16 = matmul(transpose_x = var_3009_transpose_x_0, transpose_y = var_3009_transpose_y_0, x = var_3008_cast_fp16, y = block_cache_map)[name = string("op_3009_cast_fp16")];
            tensor<int32, [1]> nk_full_13_axes_0 = const()[name = string("nk_full_13_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_13_cast_fp16 = expand_dims(axes = nk_full_13_axes_0, x = var_3009_cast_fp16)[name = string("nk_full_13_cast_fp16")];
            tensor<int32, [4]> var_3028 = const()[name = string("op_3028"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_6_cast_fp16 = transpose(perm = transpose_6_perm_0, x = v_13_cast_fp16)[name = string("transpose_147")];
            tensor<fp16, [1, 1024, 1, 16]> var_3029_cast_fp16 = reshape(shape = var_3028, x = transpose_6_cast_fp16)[name = string("op_3029_cast_fp16")];
            tensor<int32, [1]> var_3031_axes_0 = const()[name = string("op_3031_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3031_cast_fp16 = squeeze(axes = var_3031_axes_0, x = var_3029_cast_fp16)[name = string("op_3031_cast_fp16")];
            bool var_3032_transpose_x_0 = const()[name = string("op_3032_transpose_x_0"), val = bool(false)];
            bool var_3032_transpose_y_0 = const()[name = string("op_3032_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3032_cast_fp16 = matmul(transpose_x = var_3032_transpose_x_0, transpose_y = var_3032_transpose_y_0, x = var_3031_cast_fp16, y = block_cache_map)[name = string("op_3032_cast_fp16")];
            tensor<int32, [1]> nv_full_13_axes_0 = const()[name = string("nv_full_13_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_13_cast_fp16 = expand_dims(axes = nv_full_13_axes_0, x = var_3032_cast_fp16)[name = string("nv_full_13_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3038_cast_fp16 = mul(x = lkc_25_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3038_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3039_cast_fp16 = mul(x = nk_full_13_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3039_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_27_cast_fp16 = add(x = var_3038_cast_fp16, y = var_3039_cast_fp16)[name = string("lkc_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3045_cast_fp16 = mul(x = lvc_25_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3045_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3046_cast_fp16 = mul(x = nv_full_13_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3046_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_27_cast_fp16 = add(x = var_3045_cast_fp16, y = var_3046_cast_fp16)[name = string("lvc_27_cast_fp16")];
            tensor<int32, [1]> var_3050_axes_0 = const()[name = string("op_3050_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3050_cast_fp16 = squeeze(axes = var_3050_axes_0, x = lkc_27_cast_fp16)[name = string("op_3050_cast_fp16")];
            tensor<int32, [4]> var_3054 = const()[name = string("op_3054"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_13_cast_fp16 = reshape(shape = var_3054, x = var_3050_cast_fp16)[name = string("kc_13_cast_fp16")];
            tensor<int32, [1]> var_3057_axes_0 = const()[name = string("op_3057_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3057_cast_fp16 = squeeze(axes = var_3057_axes_0, x = lvc_27_cast_fp16)[name = string("op_3057_cast_fp16")];
            tensor<int32, [4]> var_3061 = const()[name = string("op_3061"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_13_cast_fp16 = reshape(shape = var_3061, x = var_3057_cast_fp16)[name = string("vc_13_cast_fp16")];
            bool var_3063_transpose_x_0 = const()[name = string("op_3063_transpose_x_0"), val = bool(false)];
            bool var_3063_transpose_y_0 = const()[name = string("op_3063_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_3063_cast_fp16 = matmul(transpose_x = var_3063_transpose_x_0, transpose_y = var_3063_transpose_y_0, x = q_41_cast_fp16, y = kc_13_cast_fp16)[name = string("op_3063_cast_fp16")];
            fp16 _inversed_scores_37_y_0_to_fp16 = const()[name = string("_inversed_scores_37_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_37_cast_fp16 = mul(x = var_3063_cast_fp16, y = _inversed_scores_37_y_0_to_fp16)[name = string("_inversed_scores_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_39_cast_fp16 = add(x = _inversed_scores_37_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_39_cast_fp16")];
            int32 var_3092 = const()[name = string("op_3092"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_3094_cast_fp16 = softmax(axis = var_3092, x = scores_39_cast_fp16)[name = string("op_3094_cast_fp16")];
            bool attn_out_25_transpose_x_1 = const()[name = string("attn_out_25_transpose_x_1"), val = bool(false)];
            bool attn_out_25_transpose_y_1 = const()[name = string("attn_out_25_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_25_cast_fp16 = matmul(transpose_x = attn_out_25_transpose_x_1, transpose_y = attn_out_25_transpose_y_1, x = var_3094_cast_fp16, y = vc_13_cast_fp16)[name = string("attn_out_25_cast_fp16")];
            tensor<int32, [4]> var_3106_perm_0 = const()[name = string("op_3106_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3108 = const()[name = string("op_3108"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_3106_cast_fp16 = transpose(perm = var_3106_perm_0, x = attn_out_25_cast_fp16)[name = string("transpose_145")];
            tensor<fp16, [1, 16, 1024]> input_53_cast_fp16 = reshape(shape = var_3108, x = var_3106_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_6_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_6_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(232839232)))];
            tensor<fp16, [1, 16, 1024]> linear_45_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_self_attn_o_proj_weight_to_fp16, x = input_53_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_103_cast_fp16 = add(x = hidden_states_95_cast_fp16, y = linear_45_cast_fp16)[name = string("hidden_states_103_cast_fp16")];
            fp16 var_3118_promoted_to_fp16 = const()[name = string("op_3118_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3124_cast_fp16 = pow(x = hidden_states_103_cast_fp16, y = var_3118_promoted_to_fp16)[name = string("op_3124_cast_fp16")];
            tensor<int32, [1]> variance_27_axes_0 = const()[name = string("variance_27_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_27_keep_dims_0 = const()[name = string("variance_27_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_27_cast_fp16 = reduce_mean(axes = variance_27_axes_0, keep_dims = variance_27_keep_dims_0, x = var_3124_cast_fp16)[name = string("variance_27_cast_fp16")];
            fp16 var_3127_to_fp16 = const()[name = string("op_3127_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3128_cast_fp16 = add(x = variance_27_cast_fp16, y = var_3127_to_fp16)[name = string("op_3128_cast_fp16")];
            fp32 var_3129_epsilon_0 = const()[name = string("op_3129_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3129_cast_fp16 = rsqrt(epsilon = var_3129_epsilon_0, x = var_3128_cast_fp16)[name = string("op_3129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_107_cast_fp16 = mul(x = hidden_states_103_cast_fp16, y = var_3129_cast_fp16)[name = string("hidden_states_107_cast_fp16")];
            tensor<fp16, [1024]> layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234936448)))];
            tensor<fp16, [1, 16, 1024]> input_55_cast_fp16 = mul(x = layers_6_post_attention_layernorm_weight_to_fp16, y = hidden_states_107_cast_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_6_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(234938560)))];
            tensor<fp16, [1, 16, 4096]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_gate_proj_weight_to_fp16, x = input_55_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_3142_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3142_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_6_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(243327232)))];
            tensor<fp16, [1, 16, 4096]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_6_mlp_up_proj_weight_to_fp16, x = input_55_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_59_cast_fp16 = mul(x = var_3142_cast_fp16, y = linear_47_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_6_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_6_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(251715904)))];
            tensor<fp16, [1, 16, 1024]> linear_48_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_6_mlp_down_proj_weight_to_fp16, x = input_59_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_111_cast_fp16 = add(x = hidden_states_103_cast_fp16, y = linear_48_cast_fp16)[name = string("hidden_states_111_cast_fp16")];
            tensor<int32, [4]> lkc_29_begin_0 = const()[name = string("lkc_29_begin_0"), val = tensor<int32, [4]>([0, 7168, 0, 0])];
            tensor<int32, [4]> lkc_29_end_0 = const()[name = string("lkc_29_end_0"), val = tensor<int32, [4]>([1, 8192, 1, 1024])];
            tensor<bool, [4]> lkc_29_end_mask_0 = const()[name = string("lkc_29_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_29_cast_fp16 = slice_by_index(begin = lkc_29_begin_0, end = lkc_29_end_0, end_mask = lkc_29_end_mask_0, x = key_cache)[name = string("lkc_29_cast_fp16")];
            tensor<int32, [4]> lvc_29_begin_0 = const()[name = string("lvc_29_begin_0"), val = tensor<int32, [4]>([0, 7168, 0, 0])];
            tensor<int32, [4]> lvc_29_end_0 = const()[name = string("lvc_29_end_0"), val = tensor<int32, [4]>([1, 8192, 1, 1024])];
            tensor<bool, [4]> lvc_29_end_mask_0 = const()[name = string("lvc_29_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_29_cast_fp16 = slice_by_index(begin = lvc_29_begin_0, end = lvc_29_end_0, end_mask = lvc_29_end_mask_0, x = value_cache)[name = string("lvc_29_cast_fp16")];
            fp16 var_3173_promoted_to_fp16 = const()[name = string("op_3173_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3179_cast_fp16 = pow(x = hidden_states_111_cast_fp16, y = var_3173_promoted_to_fp16)[name = string("op_3179_cast_fp16")];
            tensor<int32, [1]> variance_29_axes_0 = const()[name = string("variance_29_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_29_keep_dims_0 = const()[name = string("variance_29_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_29_cast_fp16 = reduce_mean(axes = variance_29_axes_0, keep_dims = variance_29_keep_dims_0, x = var_3179_cast_fp16)[name = string("variance_29_cast_fp16")];
            fp16 var_3182_to_fp16 = const()[name = string("op_3182_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3183_cast_fp16 = add(x = variance_29_cast_fp16, y = var_3182_to_fp16)[name = string("op_3183_cast_fp16")];
            fp32 var_3184_epsilon_0 = const()[name = string("op_3184_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3184_cast_fp16 = rsqrt(epsilon = var_3184_epsilon_0, x = var_3183_cast_fp16)[name = string("op_3184_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_115_cast_fp16 = mul(x = hidden_states_111_cast_fp16, y = var_3184_cast_fp16)[name = string("hidden_states_115_cast_fp16")];
            tensor<fp16, [1024]> layers_7_input_layernorm_weight_to_fp16 = const()[name = string("layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260104576)))];
            tensor<fp16, [1, 16, 1024]> hidden_29_cast_fp16 = mul(x = layers_7_input_layernorm_weight_to_fp16, y = hidden_states_115_cast_fp16)[name = string("hidden_29_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(260106688)))];
            tensor<fp16, [1, 16, 1024]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_q_proj_weight_to_fp16, x = hidden_29_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_3208 = const()[name = string("op_3208"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_43_cast_fp16 = reshape(shape = var_3208, x = linear_49_cast_fp16)[name = string("q_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(262203904)))];
            tensor<fp16, [1, 16, 1024]> linear_50_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_k_proj_weight_to_fp16, x = hidden_29_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3215 = const()[name = string("op_3215"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_29_cast_fp16 = reshape(shape = var_3215, x = linear_50_cast_fp16)[name = string("k_29_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(264301120)))];
            tensor<fp16, [1, 16, 1024]> linear_51_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_v_proj_weight_to_fp16, x = hidden_29_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3222 = const()[name = string("op_3222"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_15_cast_fp16 = reshape(shape = var_3222, x = linear_51_cast_fp16)[name = string("v_15_cast_fp16")];
            tensor<int32, [4]> q_45_perm_0 = const()[name = string("q_45_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_31_perm_0 = const()[name = string("k_31_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_7_perm_0 = const()[name = string("transpose_7_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_45_cast_fp16 = transpose(perm = q_45_perm_0, x = q_43_cast_fp16)[name = string("transpose_144")];
            tensor<fp16, [1, 16, 16, 64]> var_3235_cast_fp16 = mul(x = q_45_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3235_cast_fp16")];
            tensor<int32, [4]> x1_29_begin_0 = const()[name = string("x1_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_29_end_0 = const()[name = string("x1_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_29_end_mask_0 = const()[name = string("x1_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_29_cast_fp16 = slice_by_index(begin = x1_29_begin_0, end = x1_29_end_0, end_mask = x1_29_end_mask_0, x = q_45_cast_fp16)[name = string("x1_29_cast_fp16")];
            tensor<int32, [4]> x2_29_begin_0 = const()[name = string("x2_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_29_end_0 = const()[name = string("x2_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_29_end_mask_0 = const()[name = string("x2_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_29_cast_fp16 = slice_by_index(begin = x2_29_begin_0, end = x2_29_end_0, end_mask = x2_29_end_mask_0, x = q_45_cast_fp16)[name = string("x2_29_cast_fp16")];
            fp16 const_91_promoted_to_fp16 = const()[name = string("const_91_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3256_cast_fp16 = mul(x = x2_29_cast_fp16, y = const_91_promoted_to_fp16)[name = string("op_3256_cast_fp16")];
            int32 var_3258 = const()[name = string("op_3258"), val = int32(-1)];
            bool var_3259_interleave_0 = const()[name = string("op_3259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3259_cast_fp16 = concat(axis = var_3258, interleave = var_3259_interleave_0, values = (var_3256_cast_fp16, x1_29_cast_fp16))[name = string("op_3259_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3262_cast_fp16 = mul(x = var_3259_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3262_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_47_cast_fp16 = add(x = var_3235_cast_fp16, y = var_3262_cast_fp16)[name = string("q_47_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_31_cast_fp16 = transpose(perm = k_31_perm_0, x = k_29_cast_fp16)[name = string("transpose_143")];
            tensor<fp16, [1, 16, 16, 64]> var_3267_cast_fp16 = mul(x = k_31_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3267_cast_fp16")];
            tensor<int32, [4]> x1_31_begin_0 = const()[name = string("x1_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_31_end_0 = const()[name = string("x1_31_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_31_end_mask_0 = const()[name = string("x1_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_31_cast_fp16 = slice_by_index(begin = x1_31_begin_0, end = x1_31_end_0, end_mask = x1_31_end_mask_0, x = k_31_cast_fp16)[name = string("x1_31_cast_fp16")];
            tensor<int32, [4]> x2_31_begin_0 = const()[name = string("x2_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_31_end_0 = const()[name = string("x2_31_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_31_end_mask_0 = const()[name = string("x2_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_31_cast_fp16 = slice_by_index(begin = x2_31_begin_0, end = x2_31_end_0, end_mask = x2_31_end_mask_0, x = k_31_cast_fp16)[name = string("x2_31_cast_fp16")];
            fp16 const_94_promoted_to_fp16 = const()[name = string("const_94_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3288_cast_fp16 = mul(x = x2_31_cast_fp16, y = const_94_promoted_to_fp16)[name = string("op_3288_cast_fp16")];
            int32 var_3290 = const()[name = string("op_3290"), val = int32(-1)];
            bool var_3291_interleave_0 = const()[name = string("op_3291_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3291_cast_fp16 = concat(axis = var_3290, interleave = var_3291_interleave_0, values = (var_3288_cast_fp16, x1_31_cast_fp16))[name = string("op_3291_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3294_cast_fp16 = mul(x = var_3291_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_29_cast_fp16 = add(x = var_3267_cast_fp16, y = var_3294_cast_fp16)[name = string("x_29_cast_fp16")];
            tensor<int32, [4]> var_3301 = const()[name = string("op_3301"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3313 = const()[name = string("op_3313"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_3302_cast_fp16 = transpose(perm = var_3301, x = x_29_cast_fp16)[name = string("transpose_141")];
            tensor<fp16, [1, 1024, 1, 16]> var_3314_cast_fp16 = reshape(shape = var_3313, x = var_3302_cast_fp16)[name = string("op_3314_cast_fp16")];
            tensor<int32, [1]> var_3316_axes_0 = const()[name = string("op_3316_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3316_cast_fp16 = squeeze(axes = var_3316_axes_0, x = var_3314_cast_fp16)[name = string("op_3316_cast_fp16")];
            bool var_3317_transpose_x_0 = const()[name = string("op_3317_transpose_x_0"), val = bool(false)];
            bool var_3317_transpose_y_0 = const()[name = string("op_3317_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3317_cast_fp16 = matmul(transpose_x = var_3317_transpose_x_0, transpose_y = var_3317_transpose_y_0, x = var_3316_cast_fp16, y = block_cache_map)[name = string("op_3317_cast_fp16")];
            tensor<int32, [1]> nk_full_15_axes_0 = const()[name = string("nk_full_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_15_cast_fp16 = expand_dims(axes = nk_full_15_axes_0, x = var_3317_cast_fp16)[name = string("nk_full_15_cast_fp16")];
            tensor<int32, [4]> var_3336 = const()[name = string("op_3336"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_7_cast_fp16 = transpose(perm = transpose_7_perm_0, x = v_15_cast_fp16)[name = string("transpose_142")];
            tensor<fp16, [1, 1024, 1, 16]> var_3337_cast_fp16 = reshape(shape = var_3336, x = transpose_7_cast_fp16)[name = string("op_3337_cast_fp16")];
            tensor<int32, [1]> var_3339_axes_0 = const()[name = string("op_3339_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3339_cast_fp16 = squeeze(axes = var_3339_axes_0, x = var_3337_cast_fp16)[name = string("op_3339_cast_fp16")];
            bool var_3340_transpose_x_0 = const()[name = string("op_3340_transpose_x_0"), val = bool(false)];
            bool var_3340_transpose_y_0 = const()[name = string("op_3340_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3340_cast_fp16 = matmul(transpose_x = var_3340_transpose_x_0, transpose_y = var_3340_transpose_y_0, x = var_3339_cast_fp16, y = block_cache_map)[name = string("op_3340_cast_fp16")];
            tensor<int32, [1]> nv_full_15_axes_0 = const()[name = string("nv_full_15_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_15_cast_fp16 = expand_dims(axes = nv_full_15_axes_0, x = var_3340_cast_fp16)[name = string("nv_full_15_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3346_cast_fp16 = mul(x = lkc_29_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3346_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3347_cast_fp16 = mul(x = nk_full_15_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3347_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_31_cast_fp16 = add(x = var_3346_cast_fp16, y = var_3347_cast_fp16)[name = string("lkc_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3353_cast_fp16 = mul(x = lvc_29_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3354_cast_fp16 = mul(x = nv_full_15_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3354_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_31_cast_fp16 = add(x = var_3353_cast_fp16, y = var_3354_cast_fp16)[name = string("lvc_31_cast_fp16")];
            tensor<int32, [1]> var_3358_axes_0 = const()[name = string("op_3358_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3358_cast_fp16 = squeeze(axes = var_3358_axes_0, x = lkc_31_cast_fp16)[name = string("op_3358_cast_fp16")];
            tensor<int32, [4]> var_3362 = const()[name = string("op_3362"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_15_cast_fp16 = reshape(shape = var_3362, x = var_3358_cast_fp16)[name = string("kc_15_cast_fp16")];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3365_cast_fp16 = squeeze(axes = var_3365_axes_0, x = lvc_31_cast_fp16)[name = string("op_3365_cast_fp16")];
            tensor<int32, [4]> var_3369 = const()[name = string("op_3369"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_15_cast_fp16 = reshape(shape = var_3369, x = var_3365_cast_fp16)[name = string("vc_15_cast_fp16")];
            bool var_3371_transpose_x_0 = const()[name = string("op_3371_transpose_x_0"), val = bool(false)];
            bool var_3371_transpose_y_0 = const()[name = string("op_3371_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_3371_cast_fp16 = matmul(transpose_x = var_3371_transpose_x_0, transpose_y = var_3371_transpose_y_0, x = q_47_cast_fp16, y = kc_15_cast_fp16)[name = string("op_3371_cast_fp16")];
            fp16 _inversed_scores_43_y_0_to_fp16 = const()[name = string("_inversed_scores_43_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_43_cast_fp16 = mul(x = var_3371_cast_fp16, y = _inversed_scores_43_y_0_to_fp16)[name = string("_inversed_scores_43_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_45_cast_fp16 = add(x = _inversed_scores_43_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_45_cast_fp16")];
            int32 var_3400 = const()[name = string("op_3400"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_3402_cast_fp16 = softmax(axis = var_3400, x = scores_45_cast_fp16)[name = string("op_3402_cast_fp16")];
            bool attn_out_29_transpose_x_1 = const()[name = string("attn_out_29_transpose_x_1"), val = bool(false)];
            bool attn_out_29_transpose_y_1 = const()[name = string("attn_out_29_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_29_cast_fp16 = matmul(transpose_x = attn_out_29_transpose_x_1, transpose_y = attn_out_29_transpose_y_1, x = var_3402_cast_fp16, y = vc_15_cast_fp16)[name = string("attn_out_29_cast_fp16")];
            tensor<int32, [4]> var_3414_perm_0 = const()[name = string("op_3414_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3416 = const()[name = string("op_3416"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_3414_cast_fp16 = transpose(perm = var_3414_perm_0, x = attn_out_29_cast_fp16)[name = string("transpose_140")];
            tensor<fp16, [1, 16, 1024]> input_61_cast_fp16 = reshape(shape = var_3416, x = var_3414_cast_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_7_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_7_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(266398336)))];
            tensor<fp16, [1, 16, 1024]> linear_52_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_self_attn_o_proj_weight_to_fp16, x = input_61_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_119_cast_fp16 = add(x = hidden_states_111_cast_fp16, y = linear_52_cast_fp16)[name = string("hidden_states_119_cast_fp16")];
            fp16 var_3426_promoted_to_fp16 = const()[name = string("op_3426_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3432_cast_fp16 = pow(x = hidden_states_119_cast_fp16, y = var_3426_promoted_to_fp16)[name = string("op_3432_cast_fp16")];
            tensor<int32, [1]> variance_31_axes_0 = const()[name = string("variance_31_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_31_keep_dims_0 = const()[name = string("variance_31_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_31_cast_fp16 = reduce_mean(axes = variance_31_axes_0, keep_dims = variance_31_keep_dims_0, x = var_3432_cast_fp16)[name = string("variance_31_cast_fp16")];
            fp16 var_3435_to_fp16 = const()[name = string("op_3435_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3436_cast_fp16 = add(x = variance_31_cast_fp16, y = var_3435_to_fp16)[name = string("op_3436_cast_fp16")];
            fp32 var_3437_epsilon_0 = const()[name = string("op_3437_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3437_cast_fp16 = rsqrt(epsilon = var_3437_epsilon_0, x = var_3436_cast_fp16)[name = string("op_3437_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_123_cast_fp16 = mul(x = hidden_states_119_cast_fp16, y = var_3437_cast_fp16)[name = string("hidden_states_123_cast_fp16")];
            tensor<fp16, [1024]> layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268495552)))];
            tensor<fp16, [1, 16, 1024]> input_63_cast_fp16 = mul(x = layers_7_post_attention_layernorm_weight_to_fp16, y = hidden_states_123_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_7_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(268497664)))];
            tensor<fp16, [1, 16, 4096]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_gate_proj_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_3450_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_3450_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_7_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(276886336)))];
            tensor<fp16, [1, 16, 4096]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_7_mlp_up_proj_weight_to_fp16, x = input_63_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_67_cast_fp16 = mul(x = var_3450_cast_fp16, y = linear_54_cast_fp16)[name = string("input_67_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_7_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_7_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(285275008)))];
            tensor<fp16, [1, 16, 1024]> linear_55_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_7_mlp_down_proj_weight_to_fp16, x = input_67_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_127_cast_fp16 = add(x = hidden_states_119_cast_fp16, y = linear_55_cast_fp16)[name = string("hidden_states_127_cast_fp16")];
            tensor<int32, [4]> lkc_33_begin_0 = const()[name = string("lkc_33_begin_0"), val = tensor<int32, [4]>([0, 8192, 0, 0])];
            tensor<int32, [4]> lkc_33_end_0 = const()[name = string("lkc_33_end_0"), val = tensor<int32, [4]>([1, 9216, 1, 1024])];
            tensor<bool, [4]> lkc_33_end_mask_0 = const()[name = string("lkc_33_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_33_cast_fp16 = slice_by_index(begin = lkc_33_begin_0, end = lkc_33_end_0, end_mask = lkc_33_end_mask_0, x = key_cache)[name = string("lkc_33_cast_fp16")];
            tensor<int32, [4]> lvc_33_begin_0 = const()[name = string("lvc_33_begin_0"), val = tensor<int32, [4]>([0, 8192, 0, 0])];
            tensor<int32, [4]> lvc_33_end_0 = const()[name = string("lvc_33_end_0"), val = tensor<int32, [4]>([1, 9216, 1, 1024])];
            tensor<bool, [4]> lvc_33_end_mask_0 = const()[name = string("lvc_33_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_33_cast_fp16 = slice_by_index(begin = lvc_33_begin_0, end = lvc_33_end_0, end_mask = lvc_33_end_mask_0, x = value_cache)[name = string("lvc_33_cast_fp16")];
            fp16 var_3481_promoted_to_fp16 = const()[name = string("op_3481_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3487_cast_fp16 = pow(x = hidden_states_127_cast_fp16, y = var_3481_promoted_to_fp16)[name = string("op_3487_cast_fp16")];
            tensor<int32, [1]> variance_33_axes_0 = const()[name = string("variance_33_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_33_keep_dims_0 = const()[name = string("variance_33_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_33_cast_fp16 = reduce_mean(axes = variance_33_axes_0, keep_dims = variance_33_keep_dims_0, x = var_3487_cast_fp16)[name = string("variance_33_cast_fp16")];
            fp16 var_3490_to_fp16 = const()[name = string("op_3490_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3491_cast_fp16 = add(x = variance_33_cast_fp16, y = var_3490_to_fp16)[name = string("op_3491_cast_fp16")];
            fp32 var_3492_epsilon_0 = const()[name = string("op_3492_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3492_cast_fp16 = rsqrt(epsilon = var_3492_epsilon_0, x = var_3491_cast_fp16)[name = string("op_3492_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_131_cast_fp16 = mul(x = hidden_states_127_cast_fp16, y = var_3492_cast_fp16)[name = string("hidden_states_131_cast_fp16")];
            tensor<fp16, [1024]> layers_8_input_layernorm_weight_to_fp16 = const()[name = string("layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293663680)))];
            tensor<fp16, [1, 16, 1024]> hidden_33_cast_fp16 = mul(x = layers_8_input_layernorm_weight_to_fp16, y = hidden_states_131_cast_fp16)[name = string("hidden_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(293665792)))];
            tensor<fp16, [1, 16, 1024]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_q_proj_weight_to_fp16, x = hidden_33_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_3516 = const()[name = string("op_3516"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_49_cast_fp16 = reshape(shape = var_3516, x = linear_56_cast_fp16)[name = string("q_49_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(295763008)))];
            tensor<fp16, [1, 16, 1024]> linear_57_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_k_proj_weight_to_fp16, x = hidden_33_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_3523 = const()[name = string("op_3523"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_33_cast_fp16 = reshape(shape = var_3523, x = linear_57_cast_fp16)[name = string("k_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(297860224)))];
            tensor<fp16, [1, 16, 1024]> linear_58_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_v_proj_weight_to_fp16, x = hidden_33_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_3530 = const()[name = string("op_3530"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_17_cast_fp16 = reshape(shape = var_3530, x = linear_58_cast_fp16)[name = string("v_17_cast_fp16")];
            tensor<int32, [4]> q_51_perm_0 = const()[name = string("q_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_35_perm_0 = const()[name = string("k_35_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_8_perm_0 = const()[name = string("transpose_8_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_51_cast_fp16 = transpose(perm = q_51_perm_0, x = q_49_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 64]> var_3543_cast_fp16 = mul(x = q_51_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3543_cast_fp16")];
            tensor<int32, [4]> x1_33_begin_0 = const()[name = string("x1_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_33_end_0 = const()[name = string("x1_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_33_end_mask_0 = const()[name = string("x1_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_33_cast_fp16 = slice_by_index(begin = x1_33_begin_0, end = x1_33_end_0, end_mask = x1_33_end_mask_0, x = q_51_cast_fp16)[name = string("x1_33_cast_fp16")];
            tensor<int32, [4]> x2_33_begin_0 = const()[name = string("x2_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_33_end_0 = const()[name = string("x2_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_33_end_mask_0 = const()[name = string("x2_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_33_cast_fp16 = slice_by_index(begin = x2_33_begin_0, end = x2_33_end_0, end_mask = x2_33_end_mask_0, x = q_51_cast_fp16)[name = string("x2_33_cast_fp16")];
            fp16 const_103_promoted_to_fp16 = const()[name = string("const_103_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3564_cast_fp16 = mul(x = x2_33_cast_fp16, y = const_103_promoted_to_fp16)[name = string("op_3564_cast_fp16")];
            int32 var_3566 = const()[name = string("op_3566"), val = int32(-1)];
            bool var_3567_interleave_0 = const()[name = string("op_3567_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3567_cast_fp16 = concat(axis = var_3566, interleave = var_3567_interleave_0, values = (var_3564_cast_fp16, x1_33_cast_fp16))[name = string("op_3567_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3570_cast_fp16 = mul(x = var_3567_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3570_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_53_cast_fp16 = add(x = var_3543_cast_fp16, y = var_3570_cast_fp16)[name = string("q_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_35_cast_fp16 = transpose(perm = k_35_perm_0, x = k_33_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 16, 16, 64]> var_3575_cast_fp16 = mul(x = k_35_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3575_cast_fp16")];
            tensor<int32, [4]> x1_35_begin_0 = const()[name = string("x1_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_35_end_0 = const()[name = string("x1_35_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_35_end_mask_0 = const()[name = string("x1_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_35_cast_fp16 = slice_by_index(begin = x1_35_begin_0, end = x1_35_end_0, end_mask = x1_35_end_mask_0, x = k_35_cast_fp16)[name = string("x1_35_cast_fp16")];
            tensor<int32, [4]> x2_35_begin_0 = const()[name = string("x2_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_35_end_0 = const()[name = string("x2_35_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_35_end_mask_0 = const()[name = string("x2_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_35_cast_fp16 = slice_by_index(begin = x2_35_begin_0, end = x2_35_end_0, end_mask = x2_35_end_mask_0, x = k_35_cast_fp16)[name = string("x2_35_cast_fp16")];
            fp16 const_106_promoted_to_fp16 = const()[name = string("const_106_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3596_cast_fp16 = mul(x = x2_35_cast_fp16, y = const_106_promoted_to_fp16)[name = string("op_3596_cast_fp16")];
            int32 var_3598 = const()[name = string("op_3598"), val = int32(-1)];
            bool var_3599_interleave_0 = const()[name = string("op_3599_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3599_cast_fp16 = concat(axis = var_3598, interleave = var_3599_interleave_0, values = (var_3596_cast_fp16, x1_35_cast_fp16))[name = string("op_3599_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3602_cast_fp16 = mul(x = var_3599_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3602_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_33_cast_fp16 = add(x = var_3575_cast_fp16, y = var_3602_cast_fp16)[name = string("x_33_cast_fp16")];
            tensor<int32, [4]> var_3609 = const()[name = string("op_3609"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_3610_cast_fp16 = transpose(perm = var_3609, x = x_33_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 1024, 1, 16]> var_3622_cast_fp16 = reshape(shape = var_3621, x = var_3610_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<int32, [1]> var_3624_axes_0 = const()[name = string("op_3624_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3624_cast_fp16 = squeeze(axes = var_3624_axes_0, x = var_3622_cast_fp16)[name = string("op_3624_cast_fp16")];
            bool var_3625_transpose_x_0 = const()[name = string("op_3625_transpose_x_0"), val = bool(false)];
            bool var_3625_transpose_y_0 = const()[name = string("op_3625_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3625_cast_fp16 = matmul(transpose_x = var_3625_transpose_x_0, transpose_y = var_3625_transpose_y_0, x = var_3624_cast_fp16, y = block_cache_map)[name = string("op_3625_cast_fp16")];
            tensor<int32, [1]> nk_full_17_axes_0 = const()[name = string("nk_full_17_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_17_cast_fp16 = expand_dims(axes = nk_full_17_axes_0, x = var_3625_cast_fp16)[name = string("nk_full_17_cast_fp16")];
            tensor<int32, [4]> var_3644 = const()[name = string("op_3644"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_8_cast_fp16 = transpose(perm = transpose_8_perm_0, x = v_17_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 1024, 1, 16]> var_3645_cast_fp16 = reshape(shape = var_3644, x = transpose_8_cast_fp16)[name = string("op_3645_cast_fp16")];
            tensor<int32, [1]> var_3647_axes_0 = const()[name = string("op_3647_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3647_cast_fp16 = squeeze(axes = var_3647_axes_0, x = var_3645_cast_fp16)[name = string("op_3647_cast_fp16")];
            bool var_3648_transpose_x_0 = const()[name = string("op_3648_transpose_x_0"), val = bool(false)];
            bool var_3648_transpose_y_0 = const()[name = string("op_3648_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3648_cast_fp16 = matmul(transpose_x = var_3648_transpose_x_0, transpose_y = var_3648_transpose_y_0, x = var_3647_cast_fp16, y = block_cache_map)[name = string("op_3648_cast_fp16")];
            tensor<int32, [1]> nv_full_17_axes_0 = const()[name = string("nv_full_17_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_17_cast_fp16 = expand_dims(axes = nv_full_17_axes_0, x = var_3648_cast_fp16)[name = string("nv_full_17_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3654_cast_fp16 = mul(x = lkc_33_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3654_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3655_cast_fp16 = mul(x = nk_full_17_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3655_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_35_cast_fp16 = add(x = var_3654_cast_fp16, y = var_3655_cast_fp16)[name = string("lkc_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3661_cast_fp16 = mul(x = lvc_33_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3661_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3662_cast_fp16 = mul(x = nv_full_17_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3662_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_35_cast_fp16 = add(x = var_3661_cast_fp16, y = var_3662_cast_fp16)[name = string("lvc_35_cast_fp16")];
            tensor<int32, [1]> var_3666_axes_0 = const()[name = string("op_3666_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3666_cast_fp16 = squeeze(axes = var_3666_axes_0, x = lkc_35_cast_fp16)[name = string("op_3666_cast_fp16")];
            tensor<int32, [4]> var_3670 = const()[name = string("op_3670"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_17_cast_fp16 = reshape(shape = var_3670, x = var_3666_cast_fp16)[name = string("kc_17_cast_fp16")];
            tensor<int32, [1]> var_3673_axes_0 = const()[name = string("op_3673_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3673_cast_fp16 = squeeze(axes = var_3673_axes_0, x = lvc_35_cast_fp16)[name = string("op_3673_cast_fp16")];
            tensor<int32, [4]> var_3677 = const()[name = string("op_3677"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_17_cast_fp16 = reshape(shape = var_3677, x = var_3673_cast_fp16)[name = string("vc_17_cast_fp16")];
            bool var_3679_transpose_x_0 = const()[name = string("op_3679_transpose_x_0"), val = bool(false)];
            bool var_3679_transpose_y_0 = const()[name = string("op_3679_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_3679_cast_fp16 = matmul(transpose_x = var_3679_transpose_x_0, transpose_y = var_3679_transpose_y_0, x = q_53_cast_fp16, y = kc_17_cast_fp16)[name = string("op_3679_cast_fp16")];
            fp16 _inversed_scores_49_y_0_to_fp16 = const()[name = string("_inversed_scores_49_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_49_cast_fp16 = mul(x = var_3679_cast_fp16, y = _inversed_scores_49_y_0_to_fp16)[name = string("_inversed_scores_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_51_cast_fp16 = add(x = _inversed_scores_49_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_51_cast_fp16")];
            int32 var_3708 = const()[name = string("op_3708"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_3710_cast_fp16 = softmax(axis = var_3708, x = scores_51_cast_fp16)[name = string("op_3710_cast_fp16")];
            bool attn_out_33_transpose_x_1 = const()[name = string("attn_out_33_transpose_x_1"), val = bool(false)];
            bool attn_out_33_transpose_y_1 = const()[name = string("attn_out_33_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_33_cast_fp16 = matmul(transpose_x = attn_out_33_transpose_x_1, transpose_y = attn_out_33_transpose_y_1, x = var_3710_cast_fp16, y = vc_17_cast_fp16)[name = string("attn_out_33_cast_fp16")];
            tensor<int32, [4]> var_3722_perm_0 = const()[name = string("op_3722_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3724 = const()[name = string("op_3724"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_3722_cast_fp16 = transpose(perm = var_3722_perm_0, x = attn_out_33_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 1024]> input_69_cast_fp16 = reshape(shape = var_3724, x = var_3722_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_8_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_8_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(299957440)))];
            tensor<fp16, [1, 16, 1024]> linear_59_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_self_attn_o_proj_weight_to_fp16, x = input_69_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_135_cast_fp16 = add(x = hidden_states_127_cast_fp16, y = linear_59_cast_fp16)[name = string("hidden_states_135_cast_fp16")];
            fp16 var_3734_promoted_to_fp16 = const()[name = string("op_3734_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3740_cast_fp16 = pow(x = hidden_states_135_cast_fp16, y = var_3734_promoted_to_fp16)[name = string("op_3740_cast_fp16")];
            tensor<int32, [1]> variance_35_axes_0 = const()[name = string("variance_35_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_35_keep_dims_0 = const()[name = string("variance_35_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_35_cast_fp16 = reduce_mean(axes = variance_35_axes_0, keep_dims = variance_35_keep_dims_0, x = var_3740_cast_fp16)[name = string("variance_35_cast_fp16")];
            fp16 var_3743_to_fp16 = const()[name = string("op_3743_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3744_cast_fp16 = add(x = variance_35_cast_fp16, y = var_3743_to_fp16)[name = string("op_3744_cast_fp16")];
            fp32 var_3745_epsilon_0 = const()[name = string("op_3745_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3745_cast_fp16 = rsqrt(epsilon = var_3745_epsilon_0, x = var_3744_cast_fp16)[name = string("op_3745_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_139_cast_fp16 = mul(x = hidden_states_135_cast_fp16, y = var_3745_cast_fp16)[name = string("hidden_states_139_cast_fp16")];
            tensor<fp16, [1024]> layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302054656)))];
            tensor<fp16, [1, 16, 1024]> input_71_cast_fp16 = mul(x = layers_8_post_attention_layernorm_weight_to_fp16, y = hidden_states_139_cast_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_8_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(302056768)))];
            tensor<fp16, [1, 16, 4096]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_gate_proj_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_3758_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_3758_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_8_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(310445440)))];
            tensor<fp16, [1, 16, 4096]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_8_mlp_up_proj_weight_to_fp16, x = input_71_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_75_cast_fp16 = mul(x = var_3758_cast_fp16, y = linear_61_cast_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_8_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_8_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(318834112)))];
            tensor<fp16, [1, 16, 1024]> linear_62_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_8_mlp_down_proj_weight_to_fp16, x = input_75_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_143_cast_fp16 = add(x = hidden_states_135_cast_fp16, y = linear_62_cast_fp16)[name = string("hidden_states_143_cast_fp16")];
            tensor<int32, [4]> lkc_37_begin_0 = const()[name = string("lkc_37_begin_0"), val = tensor<int32, [4]>([0, 9216, 0, 0])];
            tensor<int32, [4]> lkc_37_end_0 = const()[name = string("lkc_37_end_0"), val = tensor<int32, [4]>([1, 10240, 1, 1024])];
            tensor<bool, [4]> lkc_37_end_mask_0 = const()[name = string("lkc_37_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_37_cast_fp16 = slice_by_index(begin = lkc_37_begin_0, end = lkc_37_end_0, end_mask = lkc_37_end_mask_0, x = key_cache)[name = string("lkc_37_cast_fp16")];
            tensor<int32, [4]> lvc_37_begin_0 = const()[name = string("lvc_37_begin_0"), val = tensor<int32, [4]>([0, 9216, 0, 0])];
            tensor<int32, [4]> lvc_37_end_0 = const()[name = string("lvc_37_end_0"), val = tensor<int32, [4]>([1, 10240, 1, 1024])];
            tensor<bool, [4]> lvc_37_end_mask_0 = const()[name = string("lvc_37_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_37_cast_fp16 = slice_by_index(begin = lvc_37_begin_0, end = lvc_37_end_0, end_mask = lvc_37_end_mask_0, x = value_cache)[name = string("lvc_37_cast_fp16")];
            fp16 var_3789_promoted_to_fp16 = const()[name = string("op_3789_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3795_cast_fp16 = pow(x = hidden_states_143_cast_fp16, y = var_3789_promoted_to_fp16)[name = string("op_3795_cast_fp16")];
            tensor<int32, [1]> variance_37_axes_0 = const()[name = string("variance_37_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_37_keep_dims_0 = const()[name = string("variance_37_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_37_cast_fp16 = reduce_mean(axes = variance_37_axes_0, keep_dims = variance_37_keep_dims_0, x = var_3795_cast_fp16)[name = string("variance_37_cast_fp16")];
            fp16 var_3798_to_fp16 = const()[name = string("op_3798_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_3799_cast_fp16 = add(x = variance_37_cast_fp16, y = var_3798_to_fp16)[name = string("op_3799_cast_fp16")];
            fp32 var_3800_epsilon_0 = const()[name = string("op_3800_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3800_cast_fp16 = rsqrt(epsilon = var_3800_epsilon_0, x = var_3799_cast_fp16)[name = string("op_3800_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_147_cast_fp16 = mul(x = hidden_states_143_cast_fp16, y = var_3800_cast_fp16)[name = string("hidden_states_147_cast_fp16")];
            tensor<fp16, [1024]> layers_9_input_layernorm_weight_to_fp16 = const()[name = string("layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327222784)))];
            tensor<fp16, [1, 16, 1024]> hidden_37_cast_fp16 = mul(x = layers_9_input_layernorm_weight_to_fp16, y = hidden_states_147_cast_fp16)[name = string("hidden_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327224896)))];
            tensor<fp16, [1, 16, 1024]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_q_proj_weight_to_fp16, x = hidden_37_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_3824 = const()[name = string("op_3824"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_55_cast_fp16 = reshape(shape = var_3824, x = linear_63_cast_fp16)[name = string("q_55_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329322112)))];
            tensor<fp16, [1, 16, 1024]> linear_64_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_k_proj_weight_to_fp16, x = hidden_37_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_3831 = const()[name = string("op_3831"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_37_cast_fp16 = reshape(shape = var_3831, x = linear_64_cast_fp16)[name = string("k_37_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331419328)))];
            tensor<fp16, [1, 16, 1024]> linear_65_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_v_proj_weight_to_fp16, x = hidden_37_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_3838 = const()[name = string("op_3838"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_19_cast_fp16 = reshape(shape = var_3838, x = linear_65_cast_fp16)[name = string("v_19_cast_fp16")];
            tensor<int32, [4]> q_57_perm_0 = const()[name = string("q_57_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_39_perm_0 = const()[name = string("k_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_9_perm_0 = const()[name = string("transpose_9_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_57_cast_fp16 = transpose(perm = q_57_perm_0, x = q_55_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 64]> var_3851_cast_fp16 = mul(x = q_57_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3851_cast_fp16")];
            tensor<int32, [4]> x1_37_begin_0 = const()[name = string("x1_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_37_end_0 = const()[name = string("x1_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_37_end_mask_0 = const()[name = string("x1_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_37_cast_fp16 = slice_by_index(begin = x1_37_begin_0, end = x1_37_end_0, end_mask = x1_37_end_mask_0, x = q_57_cast_fp16)[name = string("x1_37_cast_fp16")];
            tensor<int32, [4]> x2_37_begin_0 = const()[name = string("x2_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_37_end_0 = const()[name = string("x2_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_37_end_mask_0 = const()[name = string("x2_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_37_cast_fp16 = slice_by_index(begin = x2_37_begin_0, end = x2_37_end_0, end_mask = x2_37_end_mask_0, x = q_57_cast_fp16)[name = string("x2_37_cast_fp16")];
            fp16 const_115_promoted_to_fp16 = const()[name = string("const_115_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3872_cast_fp16 = mul(x = x2_37_cast_fp16, y = const_115_promoted_to_fp16)[name = string("op_3872_cast_fp16")];
            int32 var_3874 = const()[name = string("op_3874"), val = int32(-1)];
            bool var_3875_interleave_0 = const()[name = string("op_3875_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3875_cast_fp16 = concat(axis = var_3874, interleave = var_3875_interleave_0, values = (var_3872_cast_fp16, x1_37_cast_fp16))[name = string("op_3875_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3878_cast_fp16 = mul(x = var_3875_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3878_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_59_cast_fp16 = add(x = var_3851_cast_fp16, y = var_3878_cast_fp16)[name = string("q_59_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_39_cast_fp16 = transpose(perm = k_39_perm_0, x = k_37_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 16, 16, 64]> var_3883_cast_fp16 = mul(x = k_39_cast_fp16, y = var_1078_cast_fp16)[name = string("op_3883_cast_fp16")];
            tensor<int32, [4]> x1_39_begin_0 = const()[name = string("x1_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_39_end_0 = const()[name = string("x1_39_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_39_end_mask_0 = const()[name = string("x1_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_39_cast_fp16 = slice_by_index(begin = x1_39_begin_0, end = x1_39_end_0, end_mask = x1_39_end_mask_0, x = k_39_cast_fp16)[name = string("x1_39_cast_fp16")];
            tensor<int32, [4]> x2_39_begin_0 = const()[name = string("x2_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_39_end_0 = const()[name = string("x2_39_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_39_end_mask_0 = const()[name = string("x2_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_39_cast_fp16 = slice_by_index(begin = x2_39_begin_0, end = x2_39_end_0, end_mask = x2_39_end_mask_0, x = k_39_cast_fp16)[name = string("x2_39_cast_fp16")];
            fp16 const_118_promoted_to_fp16 = const()[name = string("const_118_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_3904_cast_fp16 = mul(x = x2_39_cast_fp16, y = const_118_promoted_to_fp16)[name = string("op_3904_cast_fp16")];
            int32 var_3906 = const()[name = string("op_3906"), val = int32(-1)];
            bool var_3907_interleave_0 = const()[name = string("op_3907_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_3907_cast_fp16 = concat(axis = var_3906, interleave = var_3907_interleave_0, values = (var_3904_cast_fp16, x1_39_cast_fp16))[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3910_cast_fp16 = mul(x = var_3907_cast_fp16, y = var_1105_cast_fp16)[name = string("op_3910_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_37_cast_fp16 = add(x = var_3883_cast_fp16, y = var_3910_cast_fp16)[name = string("x_37_cast_fp16")];
            tensor<int32, [4]> var_3917 = const()[name = string("op_3917"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_3929 = const()[name = string("op_3929"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_3918_cast_fp16 = transpose(perm = var_3917, x = x_37_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 1024, 1, 16]> var_3930_cast_fp16 = reshape(shape = var_3929, x = var_3918_cast_fp16)[name = string("op_3930_cast_fp16")];
            tensor<int32, [1]> var_3932_axes_0 = const()[name = string("op_3932_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3932_cast_fp16 = squeeze(axes = var_3932_axes_0, x = var_3930_cast_fp16)[name = string("op_3932_cast_fp16")];
            bool var_3933_transpose_x_0 = const()[name = string("op_3933_transpose_x_0"), val = bool(false)];
            bool var_3933_transpose_y_0 = const()[name = string("op_3933_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3933_cast_fp16 = matmul(transpose_x = var_3933_transpose_x_0, transpose_y = var_3933_transpose_y_0, x = var_3932_cast_fp16, y = block_cache_map)[name = string("op_3933_cast_fp16")];
            tensor<int32, [1]> nk_full_19_axes_0 = const()[name = string("nk_full_19_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_19_cast_fp16 = expand_dims(axes = nk_full_19_axes_0, x = var_3933_cast_fp16)[name = string("nk_full_19_cast_fp16")];
            tensor<int32, [4]> var_3952 = const()[name = string("op_3952"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_9_cast_fp16 = transpose(perm = transpose_9_perm_0, x = v_19_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 1024, 1, 16]> var_3953_cast_fp16 = reshape(shape = var_3952, x = transpose_9_cast_fp16)[name = string("op_3953_cast_fp16")];
            tensor<int32, [1]> var_3955_axes_0 = const()[name = string("op_3955_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_3955_cast_fp16 = squeeze(axes = var_3955_axes_0, x = var_3953_cast_fp16)[name = string("op_3955_cast_fp16")];
            bool var_3956_transpose_x_0 = const()[name = string("op_3956_transpose_x_0"), val = bool(false)];
            bool var_3956_transpose_y_0 = const()[name = string("op_3956_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_3956_cast_fp16 = matmul(transpose_x = var_3956_transpose_x_0, transpose_y = var_3956_transpose_y_0, x = var_3955_cast_fp16, y = block_cache_map)[name = string("op_3956_cast_fp16")];
            tensor<int32, [1]> nv_full_19_axes_0 = const()[name = string("nv_full_19_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_19_cast_fp16 = expand_dims(axes = nv_full_19_axes_0, x = var_3956_cast_fp16)[name = string("nv_full_19_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3962_cast_fp16 = mul(x = lkc_37_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3962_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3963_cast_fp16 = mul(x = nk_full_19_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3963_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_39_cast_fp16 = add(x = var_3962_cast_fp16, y = var_3963_cast_fp16)[name = string("lkc_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3969_cast_fp16 = mul(x = lvc_37_cast_fp16, y = var_1189_cast_fp16)[name = string("op_3969_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_3970_cast_fp16 = mul(x = nv_full_19_cast_fp16, y = update_mask_cast_fp16)[name = string("op_3970_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_39_cast_fp16 = add(x = var_3969_cast_fp16, y = var_3970_cast_fp16)[name = string("lvc_39_cast_fp16")];
            tensor<int32, [1]> var_3974_axes_0 = const()[name = string("op_3974_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3974_cast_fp16 = squeeze(axes = var_3974_axes_0, x = lkc_39_cast_fp16)[name = string("op_3974_cast_fp16")];
            tensor<int32, [4]> var_3978 = const()[name = string("op_3978"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_19_cast_fp16 = reshape(shape = var_3978, x = var_3974_cast_fp16)[name = string("kc_19_cast_fp16")];
            tensor<int32, [1]> var_3981_axes_0 = const()[name = string("op_3981_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_3981_cast_fp16 = squeeze(axes = var_3981_axes_0, x = lvc_39_cast_fp16)[name = string("op_3981_cast_fp16")];
            tensor<int32, [4]> var_3985 = const()[name = string("op_3985"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_19_cast_fp16 = reshape(shape = var_3985, x = var_3981_cast_fp16)[name = string("vc_19_cast_fp16")];
            bool var_3987_transpose_x_0 = const()[name = string("op_3987_transpose_x_0"), val = bool(false)];
            bool var_3987_transpose_y_0 = const()[name = string("op_3987_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_3987_cast_fp16 = matmul(transpose_x = var_3987_transpose_x_0, transpose_y = var_3987_transpose_y_0, x = q_59_cast_fp16, y = kc_19_cast_fp16)[name = string("op_3987_cast_fp16")];
            fp16 _inversed_scores_55_y_0_to_fp16 = const()[name = string("_inversed_scores_55_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_55_cast_fp16 = mul(x = var_3987_cast_fp16, y = _inversed_scores_55_y_0_to_fp16)[name = string("_inversed_scores_55_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_57_cast_fp16 = add(x = _inversed_scores_55_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_57_cast_fp16")];
            int32 var_4016 = const()[name = string("op_4016"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_4018_cast_fp16 = softmax(axis = var_4016, x = scores_57_cast_fp16)[name = string("op_4018_cast_fp16")];
            bool attn_out_37_transpose_x_1 = const()[name = string("attn_out_37_transpose_x_1"), val = bool(false)];
            bool attn_out_37_transpose_y_1 = const()[name = string("attn_out_37_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_37_cast_fp16 = matmul(transpose_x = attn_out_37_transpose_x_1, transpose_y = attn_out_37_transpose_y_1, x = var_4018_cast_fp16, y = vc_19_cast_fp16)[name = string("attn_out_37_cast_fp16")];
            tensor<int32, [4]> var_4030_perm_0 = const()[name = string("op_4030_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4032 = const()[name = string("op_4032"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_4030_cast_fp16 = transpose(perm = var_4030_perm_0, x = attn_out_37_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 1024]> input_77_cast_fp16 = reshape(shape = var_4032, x = var_4030_cast_fp16)[name = string("input_77_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_9_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_9_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333516544)))];
            tensor<fp16, [1, 16, 1024]> linear_66_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_self_attn_o_proj_weight_to_fp16, x = input_77_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_151_cast_fp16 = add(x = hidden_states_143_cast_fp16, y = linear_66_cast_fp16)[name = string("hidden_states_151_cast_fp16")];
            fp16 var_4042_promoted_to_fp16 = const()[name = string("op_4042_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4048_cast_fp16 = pow(x = hidden_states_151_cast_fp16, y = var_4042_promoted_to_fp16)[name = string("op_4048_cast_fp16")];
            tensor<int32, [1]> variance_39_axes_0 = const()[name = string("variance_39_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_39_keep_dims_0 = const()[name = string("variance_39_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_39_cast_fp16 = reduce_mean(axes = variance_39_axes_0, keep_dims = variance_39_keep_dims_0, x = var_4048_cast_fp16)[name = string("variance_39_cast_fp16")];
            fp16 var_4051_to_fp16 = const()[name = string("op_4051_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4052_cast_fp16 = add(x = variance_39_cast_fp16, y = var_4051_to_fp16)[name = string("op_4052_cast_fp16")];
            fp32 var_4053_epsilon_0 = const()[name = string("op_4053_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4053_cast_fp16 = rsqrt(epsilon = var_4053_epsilon_0, x = var_4052_cast_fp16)[name = string("op_4053_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_155_cast_fp16 = mul(x = hidden_states_151_cast_fp16, y = var_4053_cast_fp16)[name = string("hidden_states_155_cast_fp16")];
            tensor<fp16, [1024]> layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335613760)))];
            tensor<fp16, [1, 16, 1024]> input_79_cast_fp16 = mul(x = layers_9_post_attention_layernorm_weight_to_fp16, y = hidden_states_155_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_9_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_9_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(335615872)))];
            tensor<fp16, [1, 16, 4096]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_gate_proj_weight_to_fp16, x = input_79_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_4066_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_9_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_9_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344004544)))];
            tensor<fp16, [1, 16, 4096]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_9_mlp_up_proj_weight_to_fp16, x = input_79_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_83_cast_fp16 = mul(x = var_4066_cast_fp16, y = linear_68_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_9_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_9_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352393216)))];
            tensor<fp16, [1, 16, 1024]> linear_69_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_9_mlp_down_proj_weight_to_fp16, x = input_83_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_159_cast_fp16 = add(x = hidden_states_151_cast_fp16, y = linear_69_cast_fp16)[name = string("hidden_states_159_cast_fp16")];
            tensor<int32, [4]> lkc_41_begin_0 = const()[name = string("lkc_41_begin_0"), val = tensor<int32, [4]>([0, 10240, 0, 0])];
            tensor<int32, [4]> lkc_41_end_0 = const()[name = string("lkc_41_end_0"), val = tensor<int32, [4]>([1, 11264, 1, 1024])];
            tensor<bool, [4]> lkc_41_end_mask_0 = const()[name = string("lkc_41_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_41_cast_fp16 = slice_by_index(begin = lkc_41_begin_0, end = lkc_41_end_0, end_mask = lkc_41_end_mask_0, x = key_cache)[name = string("lkc_41_cast_fp16")];
            tensor<int32, [4]> lvc_41_begin_0 = const()[name = string("lvc_41_begin_0"), val = tensor<int32, [4]>([0, 10240, 0, 0])];
            tensor<int32, [4]> lvc_41_end_0 = const()[name = string("lvc_41_end_0"), val = tensor<int32, [4]>([1, 11264, 1, 1024])];
            tensor<bool, [4]> lvc_41_end_mask_0 = const()[name = string("lvc_41_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_41_cast_fp16 = slice_by_index(begin = lvc_41_begin_0, end = lvc_41_end_0, end_mask = lvc_41_end_mask_0, x = value_cache)[name = string("lvc_41_cast_fp16")];
            fp16 var_4097_promoted_to_fp16 = const()[name = string("op_4097_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4103_cast_fp16 = pow(x = hidden_states_159_cast_fp16, y = var_4097_promoted_to_fp16)[name = string("op_4103_cast_fp16")];
            tensor<int32, [1]> variance_41_axes_0 = const()[name = string("variance_41_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_41_keep_dims_0 = const()[name = string("variance_41_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_41_cast_fp16 = reduce_mean(axes = variance_41_axes_0, keep_dims = variance_41_keep_dims_0, x = var_4103_cast_fp16)[name = string("variance_41_cast_fp16")];
            fp16 var_4106_to_fp16 = const()[name = string("op_4106_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4107_cast_fp16 = add(x = variance_41_cast_fp16, y = var_4106_to_fp16)[name = string("op_4107_cast_fp16")];
            fp32 var_4108_epsilon_0 = const()[name = string("op_4108_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4108_cast_fp16 = rsqrt(epsilon = var_4108_epsilon_0, x = var_4107_cast_fp16)[name = string("op_4108_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_163_cast_fp16 = mul(x = hidden_states_159_cast_fp16, y = var_4108_cast_fp16)[name = string("hidden_states_163_cast_fp16")];
            tensor<fp16, [1024]> layers_10_input_layernorm_weight_to_fp16 = const()[name = string("layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360781888)))];
            tensor<fp16, [1, 16, 1024]> hidden_41_cast_fp16 = mul(x = layers_10_input_layernorm_weight_to_fp16, y = hidden_states_163_cast_fp16)[name = string("hidden_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360784000)))];
            tensor<fp16, [1, 16, 1024]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_q_proj_weight_to_fp16, x = hidden_41_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_4132 = const()[name = string("op_4132"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_61_cast_fp16 = reshape(shape = var_4132, x = linear_70_cast_fp16)[name = string("q_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362881216)))];
            tensor<fp16, [1, 16, 1024]> linear_71_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_k_proj_weight_to_fp16, x = hidden_41_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4139 = const()[name = string("op_4139"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_41_cast_fp16 = reshape(shape = var_4139, x = linear_71_cast_fp16)[name = string("k_41_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364978432)))];
            tensor<fp16, [1, 16, 1024]> linear_72_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_v_proj_weight_to_fp16, x = hidden_41_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4146 = const()[name = string("op_4146"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_21_cast_fp16 = reshape(shape = var_4146, x = linear_72_cast_fp16)[name = string("v_21_cast_fp16")];
            tensor<int32, [4]> q_63_perm_0 = const()[name = string("q_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_43_perm_0 = const()[name = string("k_43_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_10_perm_0 = const()[name = string("transpose_10_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_63_cast_fp16 = transpose(perm = q_63_perm_0, x = q_61_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 64]> var_4159_cast_fp16 = mul(x = q_63_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4159_cast_fp16")];
            tensor<int32, [4]> x1_41_begin_0 = const()[name = string("x1_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_41_end_0 = const()[name = string("x1_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_41_end_mask_0 = const()[name = string("x1_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_41_cast_fp16 = slice_by_index(begin = x1_41_begin_0, end = x1_41_end_0, end_mask = x1_41_end_mask_0, x = q_63_cast_fp16)[name = string("x1_41_cast_fp16")];
            tensor<int32, [4]> x2_41_begin_0 = const()[name = string("x2_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_41_end_0 = const()[name = string("x2_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_41_end_mask_0 = const()[name = string("x2_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_41_cast_fp16 = slice_by_index(begin = x2_41_begin_0, end = x2_41_end_0, end_mask = x2_41_end_mask_0, x = q_63_cast_fp16)[name = string("x2_41_cast_fp16")];
            fp16 const_127_promoted_to_fp16 = const()[name = string("const_127_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4180_cast_fp16 = mul(x = x2_41_cast_fp16, y = const_127_promoted_to_fp16)[name = string("op_4180_cast_fp16")];
            int32 var_4182 = const()[name = string("op_4182"), val = int32(-1)];
            bool var_4183_interleave_0 = const()[name = string("op_4183_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4183_cast_fp16 = concat(axis = var_4182, interleave = var_4183_interleave_0, values = (var_4180_cast_fp16, x1_41_cast_fp16))[name = string("op_4183_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4186_cast_fp16 = mul(x = var_4183_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4186_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_65_cast_fp16 = add(x = var_4159_cast_fp16, y = var_4186_cast_fp16)[name = string("q_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_43_cast_fp16 = transpose(perm = k_43_perm_0, x = k_41_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 16, 16, 64]> var_4191_cast_fp16 = mul(x = k_43_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4191_cast_fp16")];
            tensor<int32, [4]> x1_43_begin_0 = const()[name = string("x1_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_43_end_0 = const()[name = string("x1_43_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_43_end_mask_0 = const()[name = string("x1_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_43_cast_fp16 = slice_by_index(begin = x1_43_begin_0, end = x1_43_end_0, end_mask = x1_43_end_mask_0, x = k_43_cast_fp16)[name = string("x1_43_cast_fp16")];
            tensor<int32, [4]> x2_43_begin_0 = const()[name = string("x2_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_43_end_0 = const()[name = string("x2_43_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_43_end_mask_0 = const()[name = string("x2_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_43_cast_fp16 = slice_by_index(begin = x2_43_begin_0, end = x2_43_end_0, end_mask = x2_43_end_mask_0, x = k_43_cast_fp16)[name = string("x2_43_cast_fp16")];
            fp16 const_130_promoted_to_fp16 = const()[name = string("const_130_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4212_cast_fp16 = mul(x = x2_43_cast_fp16, y = const_130_promoted_to_fp16)[name = string("op_4212_cast_fp16")];
            int32 var_4214 = const()[name = string("op_4214"), val = int32(-1)];
            bool var_4215_interleave_0 = const()[name = string("op_4215_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4215_cast_fp16 = concat(axis = var_4214, interleave = var_4215_interleave_0, values = (var_4212_cast_fp16, x1_43_cast_fp16))[name = string("op_4215_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4218_cast_fp16 = mul(x = var_4215_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4218_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_41_cast_fp16 = add(x = var_4191_cast_fp16, y = var_4218_cast_fp16)[name = string("x_41_cast_fp16")];
            tensor<int32, [4]> var_4225 = const()[name = string("op_4225"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4237 = const()[name = string("op_4237"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_4226_cast_fp16 = transpose(perm = var_4225, x = x_41_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 1024, 1, 16]> var_4238_cast_fp16 = reshape(shape = var_4237, x = var_4226_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<int32, [1]> var_4240_axes_0 = const()[name = string("op_4240_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4240_cast_fp16 = squeeze(axes = var_4240_axes_0, x = var_4238_cast_fp16)[name = string("op_4240_cast_fp16")];
            bool var_4241_transpose_x_0 = const()[name = string("op_4241_transpose_x_0"), val = bool(false)];
            bool var_4241_transpose_y_0 = const()[name = string("op_4241_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4241_cast_fp16 = matmul(transpose_x = var_4241_transpose_x_0, transpose_y = var_4241_transpose_y_0, x = var_4240_cast_fp16, y = block_cache_map)[name = string("op_4241_cast_fp16")];
            tensor<int32, [1]> nk_full_21_axes_0 = const()[name = string("nk_full_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_21_cast_fp16 = expand_dims(axes = nk_full_21_axes_0, x = var_4241_cast_fp16)[name = string("nk_full_21_cast_fp16")];
            tensor<int32, [4]> var_4260 = const()[name = string("op_4260"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_10_cast_fp16 = transpose(perm = transpose_10_perm_0, x = v_21_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 1024, 1, 16]> var_4261_cast_fp16 = reshape(shape = var_4260, x = transpose_10_cast_fp16)[name = string("op_4261_cast_fp16")];
            tensor<int32, [1]> var_4263_axes_0 = const()[name = string("op_4263_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4263_cast_fp16 = squeeze(axes = var_4263_axes_0, x = var_4261_cast_fp16)[name = string("op_4263_cast_fp16")];
            bool var_4264_transpose_x_0 = const()[name = string("op_4264_transpose_x_0"), val = bool(false)];
            bool var_4264_transpose_y_0 = const()[name = string("op_4264_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4264_cast_fp16 = matmul(transpose_x = var_4264_transpose_x_0, transpose_y = var_4264_transpose_y_0, x = var_4263_cast_fp16, y = block_cache_map)[name = string("op_4264_cast_fp16")];
            tensor<int32, [1]> nv_full_21_axes_0 = const()[name = string("nv_full_21_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_21_cast_fp16 = expand_dims(axes = nv_full_21_axes_0, x = var_4264_cast_fp16)[name = string("nv_full_21_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4270_cast_fp16 = mul(x = lkc_41_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4270_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4271_cast_fp16 = mul(x = nk_full_21_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4271_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_43_cast_fp16 = add(x = var_4270_cast_fp16, y = var_4271_cast_fp16)[name = string("lkc_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4277_cast_fp16 = mul(x = lvc_41_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4277_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4278_cast_fp16 = mul(x = nv_full_21_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4278_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_43_cast_fp16 = add(x = var_4277_cast_fp16, y = var_4278_cast_fp16)[name = string("lvc_43_cast_fp16")];
            tensor<int32, [1]> var_4282_axes_0 = const()[name = string("op_4282_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4282_cast_fp16 = squeeze(axes = var_4282_axes_0, x = lkc_43_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<int32, [4]> var_4286 = const()[name = string("op_4286"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_21_cast_fp16 = reshape(shape = var_4286, x = var_4282_cast_fp16)[name = string("kc_21_cast_fp16")];
            tensor<int32, [1]> var_4289_axes_0 = const()[name = string("op_4289_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4289_cast_fp16 = squeeze(axes = var_4289_axes_0, x = lvc_43_cast_fp16)[name = string("op_4289_cast_fp16")];
            tensor<int32, [4]> var_4293 = const()[name = string("op_4293"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_21_cast_fp16 = reshape(shape = var_4293, x = var_4289_cast_fp16)[name = string("vc_21_cast_fp16")];
            bool var_4295_transpose_x_0 = const()[name = string("op_4295_transpose_x_0"), val = bool(false)];
            bool var_4295_transpose_y_0 = const()[name = string("op_4295_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_4295_cast_fp16 = matmul(transpose_x = var_4295_transpose_x_0, transpose_y = var_4295_transpose_y_0, x = q_65_cast_fp16, y = kc_21_cast_fp16)[name = string("op_4295_cast_fp16")];
            fp16 _inversed_scores_61_y_0_to_fp16 = const()[name = string("_inversed_scores_61_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_61_cast_fp16 = mul(x = var_4295_cast_fp16, y = _inversed_scores_61_y_0_to_fp16)[name = string("_inversed_scores_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_63_cast_fp16 = add(x = _inversed_scores_61_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_63_cast_fp16")];
            int32 var_4324 = const()[name = string("op_4324"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_4326_cast_fp16 = softmax(axis = var_4324, x = scores_63_cast_fp16)[name = string("op_4326_cast_fp16")];
            bool attn_out_41_transpose_x_1 = const()[name = string("attn_out_41_transpose_x_1"), val = bool(false)];
            bool attn_out_41_transpose_y_1 = const()[name = string("attn_out_41_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_41_cast_fp16 = matmul(transpose_x = attn_out_41_transpose_x_1, transpose_y = attn_out_41_transpose_y_1, x = var_4326_cast_fp16, y = vc_21_cast_fp16)[name = string("attn_out_41_cast_fp16")];
            tensor<int32, [4]> var_4338_perm_0 = const()[name = string("op_4338_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4340 = const()[name = string("op_4340"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_4338_cast_fp16 = transpose(perm = var_4338_perm_0, x = attn_out_41_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 1024]> input_85_cast_fp16 = reshape(shape = var_4340, x = var_4338_cast_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_10_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_10_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367075648)))];
            tensor<fp16, [1, 16, 1024]> linear_73_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_self_attn_o_proj_weight_to_fp16, x = input_85_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_167_cast_fp16 = add(x = hidden_states_159_cast_fp16, y = linear_73_cast_fp16)[name = string("hidden_states_167_cast_fp16")];
            fp16 var_4350_promoted_to_fp16 = const()[name = string("op_4350_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4356_cast_fp16 = pow(x = hidden_states_167_cast_fp16, y = var_4350_promoted_to_fp16)[name = string("op_4356_cast_fp16")];
            tensor<int32, [1]> variance_43_axes_0 = const()[name = string("variance_43_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_43_keep_dims_0 = const()[name = string("variance_43_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_43_cast_fp16 = reduce_mean(axes = variance_43_axes_0, keep_dims = variance_43_keep_dims_0, x = var_4356_cast_fp16)[name = string("variance_43_cast_fp16")];
            fp16 var_4359_to_fp16 = const()[name = string("op_4359_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4360_cast_fp16 = add(x = variance_43_cast_fp16, y = var_4359_to_fp16)[name = string("op_4360_cast_fp16")];
            fp32 var_4361_epsilon_0 = const()[name = string("op_4361_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4361_cast_fp16 = rsqrt(epsilon = var_4361_epsilon_0, x = var_4360_cast_fp16)[name = string("op_4361_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_171_cast_fp16 = mul(x = hidden_states_167_cast_fp16, y = var_4361_cast_fp16)[name = string("hidden_states_171_cast_fp16")];
            tensor<fp16, [1024]> layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369172864)))];
            tensor<fp16, [1, 16, 1024]> input_87_cast_fp16 = mul(x = layers_10_post_attention_layernorm_weight_to_fp16, y = hidden_states_171_cast_fp16)[name = string("input_87_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_10_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_10_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(369174976)))];
            tensor<fp16, [1, 16, 4096]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_gate_proj_weight_to_fp16, x = input_87_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_4374_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_4374_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_10_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_10_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377563648)))];
            tensor<fp16, [1, 16, 4096]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_10_mlp_up_proj_weight_to_fp16, x = input_87_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_91_cast_fp16 = mul(x = var_4374_cast_fp16, y = linear_75_cast_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_10_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_10_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(385952320)))];
            tensor<fp16, [1, 16, 1024]> linear_76_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_10_mlp_down_proj_weight_to_fp16, x = input_91_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_175_cast_fp16 = add(x = hidden_states_167_cast_fp16, y = linear_76_cast_fp16)[name = string("hidden_states_175_cast_fp16")];
            tensor<int32, [4]> lkc_45_begin_0 = const()[name = string("lkc_45_begin_0"), val = tensor<int32, [4]>([0, 11264, 0, 0])];
            tensor<int32, [4]> lkc_45_end_0 = const()[name = string("lkc_45_end_0"), val = tensor<int32, [4]>([1, 12288, 1, 1024])];
            tensor<bool, [4]> lkc_45_end_mask_0 = const()[name = string("lkc_45_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_45_cast_fp16 = slice_by_index(begin = lkc_45_begin_0, end = lkc_45_end_0, end_mask = lkc_45_end_mask_0, x = key_cache)[name = string("lkc_45_cast_fp16")];
            tensor<int32, [4]> lvc_45_begin_0 = const()[name = string("lvc_45_begin_0"), val = tensor<int32, [4]>([0, 11264, 0, 0])];
            tensor<int32, [4]> lvc_45_end_0 = const()[name = string("lvc_45_end_0"), val = tensor<int32, [4]>([1, 12288, 1, 1024])];
            tensor<bool, [4]> lvc_45_end_mask_0 = const()[name = string("lvc_45_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_45_cast_fp16 = slice_by_index(begin = lvc_45_begin_0, end = lvc_45_end_0, end_mask = lvc_45_end_mask_0, x = value_cache)[name = string("lvc_45_cast_fp16")];
            fp16 var_4405_promoted_to_fp16 = const()[name = string("op_4405_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4411_cast_fp16 = pow(x = hidden_states_175_cast_fp16, y = var_4405_promoted_to_fp16)[name = string("op_4411_cast_fp16")];
            tensor<int32, [1]> variance_45_axes_0 = const()[name = string("variance_45_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_45_keep_dims_0 = const()[name = string("variance_45_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_45_cast_fp16 = reduce_mean(axes = variance_45_axes_0, keep_dims = variance_45_keep_dims_0, x = var_4411_cast_fp16)[name = string("variance_45_cast_fp16")];
            fp16 var_4414_to_fp16 = const()[name = string("op_4414_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4415_cast_fp16 = add(x = variance_45_cast_fp16, y = var_4414_to_fp16)[name = string("op_4415_cast_fp16")];
            fp32 var_4416_epsilon_0 = const()[name = string("op_4416_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4416_cast_fp16 = rsqrt(epsilon = var_4416_epsilon_0, x = var_4415_cast_fp16)[name = string("op_4416_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_179_cast_fp16 = mul(x = hidden_states_175_cast_fp16, y = var_4416_cast_fp16)[name = string("hidden_states_179_cast_fp16")];
            tensor<fp16, [1024]> layers_11_input_layernorm_weight_to_fp16 = const()[name = string("layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394340992)))];
            tensor<fp16, [1, 16, 1024]> hidden_45_cast_fp16 = mul(x = layers_11_input_layernorm_weight_to_fp16, y = hidden_states_179_cast_fp16)[name = string("hidden_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394343104)))];
            tensor<fp16, [1, 16, 1024]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_q_proj_weight_to_fp16, x = hidden_45_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_4440 = const()[name = string("op_4440"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_67_cast_fp16 = reshape(shape = var_4440, x = linear_77_cast_fp16)[name = string("q_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396440320)))];
            tensor<fp16, [1, 16, 1024]> linear_78_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_k_proj_weight_to_fp16, x = hidden_45_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_4447 = const()[name = string("op_4447"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_45_cast_fp16 = reshape(shape = var_4447, x = linear_78_cast_fp16)[name = string("k_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(398537536)))];
            tensor<fp16, [1, 16, 1024]> linear_79_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_v_proj_weight_to_fp16, x = hidden_45_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_4454 = const()[name = string("op_4454"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_23_cast_fp16 = reshape(shape = var_4454, x = linear_79_cast_fp16)[name = string("v_23_cast_fp16")];
            tensor<int32, [4]> q_69_perm_0 = const()[name = string("q_69_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_47_perm_0 = const()[name = string("k_47_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_11_perm_0 = const()[name = string("transpose_11_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_69_cast_fp16 = transpose(perm = q_69_perm_0, x = q_67_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 64]> var_4467_cast_fp16 = mul(x = q_69_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4467_cast_fp16")];
            tensor<int32, [4]> x1_45_begin_0 = const()[name = string("x1_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_45_end_0 = const()[name = string("x1_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_45_end_mask_0 = const()[name = string("x1_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_45_cast_fp16 = slice_by_index(begin = x1_45_begin_0, end = x1_45_end_0, end_mask = x1_45_end_mask_0, x = q_69_cast_fp16)[name = string("x1_45_cast_fp16")];
            tensor<int32, [4]> x2_45_begin_0 = const()[name = string("x2_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_45_end_0 = const()[name = string("x2_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_45_end_mask_0 = const()[name = string("x2_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_45_cast_fp16 = slice_by_index(begin = x2_45_begin_0, end = x2_45_end_0, end_mask = x2_45_end_mask_0, x = q_69_cast_fp16)[name = string("x2_45_cast_fp16")];
            fp16 const_139_promoted_to_fp16 = const()[name = string("const_139_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4488_cast_fp16 = mul(x = x2_45_cast_fp16, y = const_139_promoted_to_fp16)[name = string("op_4488_cast_fp16")];
            int32 var_4490 = const()[name = string("op_4490"), val = int32(-1)];
            bool var_4491_interleave_0 = const()[name = string("op_4491_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4491_cast_fp16 = concat(axis = var_4490, interleave = var_4491_interleave_0, values = (var_4488_cast_fp16, x1_45_cast_fp16))[name = string("op_4491_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4494_cast_fp16 = mul(x = var_4491_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4494_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_71_cast_fp16 = add(x = var_4467_cast_fp16, y = var_4494_cast_fp16)[name = string("q_71_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_47_cast_fp16 = transpose(perm = k_47_perm_0, x = k_45_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 16, 16, 64]> var_4499_cast_fp16 = mul(x = k_47_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4499_cast_fp16")];
            tensor<int32, [4]> x1_47_begin_0 = const()[name = string("x1_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_47_end_0 = const()[name = string("x1_47_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_47_end_mask_0 = const()[name = string("x1_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_47_cast_fp16 = slice_by_index(begin = x1_47_begin_0, end = x1_47_end_0, end_mask = x1_47_end_mask_0, x = k_47_cast_fp16)[name = string("x1_47_cast_fp16")];
            tensor<int32, [4]> x2_47_begin_0 = const()[name = string("x2_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_47_end_0 = const()[name = string("x2_47_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_47_end_mask_0 = const()[name = string("x2_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_47_cast_fp16 = slice_by_index(begin = x2_47_begin_0, end = x2_47_end_0, end_mask = x2_47_end_mask_0, x = k_47_cast_fp16)[name = string("x2_47_cast_fp16")];
            fp16 const_142_promoted_to_fp16 = const()[name = string("const_142_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4520_cast_fp16 = mul(x = x2_47_cast_fp16, y = const_142_promoted_to_fp16)[name = string("op_4520_cast_fp16")];
            int32 var_4522 = const()[name = string("op_4522"), val = int32(-1)];
            bool var_4523_interleave_0 = const()[name = string("op_4523_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4523_cast_fp16 = concat(axis = var_4522, interleave = var_4523_interleave_0, values = (var_4520_cast_fp16, x1_47_cast_fp16))[name = string("op_4523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4526_cast_fp16 = mul(x = var_4523_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4526_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_45_cast_fp16 = add(x = var_4499_cast_fp16, y = var_4526_cast_fp16)[name = string("x_45_cast_fp16")];
            tensor<int32, [4]> var_4533 = const()[name = string("op_4533"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4545 = const()[name = string("op_4545"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_4534_cast_fp16 = transpose(perm = var_4533, x = x_45_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 1024, 1, 16]> var_4546_cast_fp16 = reshape(shape = var_4545, x = var_4534_cast_fp16)[name = string("op_4546_cast_fp16")];
            tensor<int32, [1]> var_4548_axes_0 = const()[name = string("op_4548_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4548_cast_fp16 = squeeze(axes = var_4548_axes_0, x = var_4546_cast_fp16)[name = string("op_4548_cast_fp16")];
            bool var_4549_transpose_x_0 = const()[name = string("op_4549_transpose_x_0"), val = bool(false)];
            bool var_4549_transpose_y_0 = const()[name = string("op_4549_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4549_cast_fp16 = matmul(transpose_x = var_4549_transpose_x_0, transpose_y = var_4549_transpose_y_0, x = var_4548_cast_fp16, y = block_cache_map)[name = string("op_4549_cast_fp16")];
            tensor<int32, [1]> nk_full_23_axes_0 = const()[name = string("nk_full_23_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_23_cast_fp16 = expand_dims(axes = nk_full_23_axes_0, x = var_4549_cast_fp16)[name = string("nk_full_23_cast_fp16")];
            tensor<int32, [4]> var_4568 = const()[name = string("op_4568"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_11_cast_fp16 = transpose(perm = transpose_11_perm_0, x = v_23_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 1024, 1, 16]> var_4569_cast_fp16 = reshape(shape = var_4568, x = transpose_11_cast_fp16)[name = string("op_4569_cast_fp16")];
            tensor<int32, [1]> var_4571_axes_0 = const()[name = string("op_4571_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4571_cast_fp16 = squeeze(axes = var_4571_axes_0, x = var_4569_cast_fp16)[name = string("op_4571_cast_fp16")];
            bool var_4572_transpose_x_0 = const()[name = string("op_4572_transpose_x_0"), val = bool(false)];
            bool var_4572_transpose_y_0 = const()[name = string("op_4572_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4572_cast_fp16 = matmul(transpose_x = var_4572_transpose_x_0, transpose_y = var_4572_transpose_y_0, x = var_4571_cast_fp16, y = block_cache_map)[name = string("op_4572_cast_fp16")];
            tensor<int32, [1]> nv_full_23_axes_0 = const()[name = string("nv_full_23_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_23_cast_fp16 = expand_dims(axes = nv_full_23_axes_0, x = var_4572_cast_fp16)[name = string("nv_full_23_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4578_cast_fp16 = mul(x = lkc_45_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4578_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4579_cast_fp16 = mul(x = nk_full_23_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4579_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_47_cast_fp16 = add(x = var_4578_cast_fp16, y = var_4579_cast_fp16)[name = string("lkc_47_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4585_cast_fp16 = mul(x = lvc_45_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4585_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4586_cast_fp16 = mul(x = nv_full_23_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4586_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_47_cast_fp16 = add(x = var_4585_cast_fp16, y = var_4586_cast_fp16)[name = string("lvc_47_cast_fp16")];
            tensor<int32, [1]> var_4590_axes_0 = const()[name = string("op_4590_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4590_cast_fp16 = squeeze(axes = var_4590_axes_0, x = lkc_47_cast_fp16)[name = string("op_4590_cast_fp16")];
            tensor<int32, [4]> var_4594 = const()[name = string("op_4594"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_23_cast_fp16 = reshape(shape = var_4594, x = var_4590_cast_fp16)[name = string("kc_23_cast_fp16")];
            tensor<int32, [1]> var_4597_axes_0 = const()[name = string("op_4597_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4597_cast_fp16 = squeeze(axes = var_4597_axes_0, x = lvc_47_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<int32, [4]> var_4601 = const()[name = string("op_4601"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_23_cast_fp16 = reshape(shape = var_4601, x = var_4597_cast_fp16)[name = string("vc_23_cast_fp16")];
            bool var_4603_transpose_x_0 = const()[name = string("op_4603_transpose_x_0"), val = bool(false)];
            bool var_4603_transpose_y_0 = const()[name = string("op_4603_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_4603_cast_fp16 = matmul(transpose_x = var_4603_transpose_x_0, transpose_y = var_4603_transpose_y_0, x = q_71_cast_fp16, y = kc_23_cast_fp16)[name = string("op_4603_cast_fp16")];
            fp16 _inversed_scores_67_y_0_to_fp16 = const()[name = string("_inversed_scores_67_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_67_cast_fp16 = mul(x = var_4603_cast_fp16, y = _inversed_scores_67_y_0_to_fp16)[name = string("_inversed_scores_67_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_69_cast_fp16 = add(x = _inversed_scores_67_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_69_cast_fp16")];
            int32 var_4632 = const()[name = string("op_4632"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_4634_cast_fp16 = softmax(axis = var_4632, x = scores_69_cast_fp16)[name = string("op_4634_cast_fp16")];
            bool attn_out_45_transpose_x_1 = const()[name = string("attn_out_45_transpose_x_1"), val = bool(false)];
            bool attn_out_45_transpose_y_1 = const()[name = string("attn_out_45_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_45_cast_fp16 = matmul(transpose_x = attn_out_45_transpose_x_1, transpose_y = attn_out_45_transpose_y_1, x = var_4634_cast_fp16, y = vc_23_cast_fp16)[name = string("attn_out_45_cast_fp16")];
            tensor<int32, [4]> var_4646_perm_0 = const()[name = string("op_4646_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4648 = const()[name = string("op_4648"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_4646_cast_fp16 = transpose(perm = var_4646_perm_0, x = attn_out_45_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 1024]> input_93_cast_fp16 = reshape(shape = var_4648, x = var_4646_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_11_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_11_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(400634752)))];
            tensor<fp16, [1, 16, 1024]> linear_80_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_self_attn_o_proj_weight_to_fp16, x = input_93_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_183_cast_fp16 = add(x = hidden_states_175_cast_fp16, y = linear_80_cast_fp16)[name = string("hidden_states_183_cast_fp16")];
            fp16 var_4658_promoted_to_fp16 = const()[name = string("op_4658_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4664_cast_fp16 = pow(x = hidden_states_183_cast_fp16, y = var_4658_promoted_to_fp16)[name = string("op_4664_cast_fp16")];
            tensor<int32, [1]> variance_47_axes_0 = const()[name = string("variance_47_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_47_keep_dims_0 = const()[name = string("variance_47_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_47_cast_fp16 = reduce_mean(axes = variance_47_axes_0, keep_dims = variance_47_keep_dims_0, x = var_4664_cast_fp16)[name = string("variance_47_cast_fp16")];
            fp16 var_4667_to_fp16 = const()[name = string("op_4667_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4668_cast_fp16 = add(x = variance_47_cast_fp16, y = var_4667_to_fp16)[name = string("op_4668_cast_fp16")];
            fp32 var_4669_epsilon_0 = const()[name = string("op_4669_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4669_cast_fp16 = rsqrt(epsilon = var_4669_epsilon_0, x = var_4668_cast_fp16)[name = string("op_4669_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_187_cast_fp16 = mul(x = hidden_states_183_cast_fp16, y = var_4669_cast_fp16)[name = string("hidden_states_187_cast_fp16")];
            tensor<fp16, [1024]> layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402731968)))];
            tensor<fp16, [1, 16, 1024]> input_95_cast_fp16 = mul(x = layers_11_post_attention_layernorm_weight_to_fp16, y = hidden_states_187_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_11_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402734080)))];
            tensor<fp16, [1, 16, 4096]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_gate_proj_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_4682_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_4682_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_11_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(411122752)))];
            tensor<fp16, [1, 16, 4096]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_11_mlp_up_proj_weight_to_fp16, x = input_95_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_99_cast_fp16 = mul(x = var_4682_cast_fp16, y = linear_82_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_11_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_11_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(419511424)))];
            tensor<fp16, [1, 16, 1024]> linear_83_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_11_mlp_down_proj_weight_to_fp16, x = input_99_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_191_cast_fp16 = add(x = hidden_states_183_cast_fp16, y = linear_83_cast_fp16)[name = string("hidden_states_191_cast_fp16")];
            tensor<int32, [4]> lkc_49_begin_0 = const()[name = string("lkc_49_begin_0"), val = tensor<int32, [4]>([0, 12288, 0, 0])];
            tensor<int32, [4]> lkc_49_end_0 = const()[name = string("lkc_49_end_0"), val = tensor<int32, [4]>([1, 13312, 1, 1024])];
            tensor<bool, [4]> lkc_49_end_mask_0 = const()[name = string("lkc_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_49_cast_fp16 = slice_by_index(begin = lkc_49_begin_0, end = lkc_49_end_0, end_mask = lkc_49_end_mask_0, x = key_cache)[name = string("lkc_49_cast_fp16")];
            tensor<int32, [4]> lvc_49_begin_0 = const()[name = string("lvc_49_begin_0"), val = tensor<int32, [4]>([0, 12288, 0, 0])];
            tensor<int32, [4]> lvc_49_end_0 = const()[name = string("lvc_49_end_0"), val = tensor<int32, [4]>([1, 13312, 1, 1024])];
            tensor<bool, [4]> lvc_49_end_mask_0 = const()[name = string("lvc_49_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_49_cast_fp16 = slice_by_index(begin = lvc_49_begin_0, end = lvc_49_end_0, end_mask = lvc_49_end_mask_0, x = value_cache)[name = string("lvc_49_cast_fp16")];
            fp16 var_4713_promoted_to_fp16 = const()[name = string("op_4713_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4719_cast_fp16 = pow(x = hidden_states_191_cast_fp16, y = var_4713_promoted_to_fp16)[name = string("op_4719_cast_fp16")];
            tensor<int32, [1]> variance_49_axes_0 = const()[name = string("variance_49_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_49_keep_dims_0 = const()[name = string("variance_49_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_49_cast_fp16 = reduce_mean(axes = variance_49_axes_0, keep_dims = variance_49_keep_dims_0, x = var_4719_cast_fp16)[name = string("variance_49_cast_fp16")];
            fp16 var_4722_to_fp16 = const()[name = string("op_4722_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4723_cast_fp16 = add(x = variance_49_cast_fp16, y = var_4722_to_fp16)[name = string("op_4723_cast_fp16")];
            fp32 var_4724_epsilon_0 = const()[name = string("op_4724_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4724_cast_fp16 = rsqrt(epsilon = var_4724_epsilon_0, x = var_4723_cast_fp16)[name = string("op_4724_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_195_cast_fp16 = mul(x = hidden_states_191_cast_fp16, y = var_4724_cast_fp16)[name = string("hidden_states_195_cast_fp16")];
            tensor<fp16, [1024]> layers_12_input_layernorm_weight_to_fp16 = const()[name = string("layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427900096)))];
            tensor<fp16, [1, 16, 1024]> hidden_49_cast_fp16 = mul(x = layers_12_input_layernorm_weight_to_fp16, y = hidden_states_195_cast_fp16)[name = string("hidden_49_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427902208)))];
            tensor<fp16, [1, 16, 1024]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_q_proj_weight_to_fp16, x = hidden_49_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_4748 = const()[name = string("op_4748"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_73_cast_fp16 = reshape(shape = var_4748, x = linear_84_cast_fp16)[name = string("q_73_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(429999424)))];
            tensor<fp16, [1, 16, 1024]> linear_85_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_k_proj_weight_to_fp16, x = hidden_49_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_4755 = const()[name = string("op_4755"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_49_cast_fp16 = reshape(shape = var_4755, x = linear_85_cast_fp16)[name = string("k_49_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(432096640)))];
            tensor<fp16, [1, 16, 1024]> linear_86_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_v_proj_weight_to_fp16, x = hidden_49_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_4762 = const()[name = string("op_4762"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_25_cast_fp16 = reshape(shape = var_4762, x = linear_86_cast_fp16)[name = string("v_25_cast_fp16")];
            tensor<int32, [4]> q_75_perm_0 = const()[name = string("q_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_51_perm_0 = const()[name = string("k_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_12_perm_0 = const()[name = string("transpose_12_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_75_cast_fp16 = transpose(perm = q_75_perm_0, x = q_73_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 64]> var_4775_cast_fp16 = mul(x = q_75_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4775_cast_fp16")];
            tensor<int32, [4]> x1_49_begin_0 = const()[name = string("x1_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_49_end_0 = const()[name = string("x1_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_49_end_mask_0 = const()[name = string("x1_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_49_cast_fp16 = slice_by_index(begin = x1_49_begin_0, end = x1_49_end_0, end_mask = x1_49_end_mask_0, x = q_75_cast_fp16)[name = string("x1_49_cast_fp16")];
            tensor<int32, [4]> x2_49_begin_0 = const()[name = string("x2_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_49_end_0 = const()[name = string("x2_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_49_end_mask_0 = const()[name = string("x2_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_49_cast_fp16 = slice_by_index(begin = x2_49_begin_0, end = x2_49_end_0, end_mask = x2_49_end_mask_0, x = q_75_cast_fp16)[name = string("x2_49_cast_fp16")];
            fp16 const_151_promoted_to_fp16 = const()[name = string("const_151_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4796_cast_fp16 = mul(x = x2_49_cast_fp16, y = const_151_promoted_to_fp16)[name = string("op_4796_cast_fp16")];
            int32 var_4798 = const()[name = string("op_4798"), val = int32(-1)];
            bool var_4799_interleave_0 = const()[name = string("op_4799_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4799_cast_fp16 = concat(axis = var_4798, interleave = var_4799_interleave_0, values = (var_4796_cast_fp16, x1_49_cast_fp16))[name = string("op_4799_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4802_cast_fp16 = mul(x = var_4799_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_77_cast_fp16 = add(x = var_4775_cast_fp16, y = var_4802_cast_fp16)[name = string("q_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_51_cast_fp16 = transpose(perm = k_51_perm_0, x = k_49_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 16, 16, 64]> var_4807_cast_fp16 = mul(x = k_51_cast_fp16, y = var_1078_cast_fp16)[name = string("op_4807_cast_fp16")];
            tensor<int32, [4]> x1_51_begin_0 = const()[name = string("x1_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_51_end_0 = const()[name = string("x1_51_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_51_end_mask_0 = const()[name = string("x1_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_51_cast_fp16 = slice_by_index(begin = x1_51_begin_0, end = x1_51_end_0, end_mask = x1_51_end_mask_0, x = k_51_cast_fp16)[name = string("x1_51_cast_fp16")];
            tensor<int32, [4]> x2_51_begin_0 = const()[name = string("x2_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_51_end_0 = const()[name = string("x2_51_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_51_end_mask_0 = const()[name = string("x2_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_51_cast_fp16 = slice_by_index(begin = x2_51_begin_0, end = x2_51_end_0, end_mask = x2_51_end_mask_0, x = k_51_cast_fp16)[name = string("x2_51_cast_fp16")];
            fp16 const_154_promoted_to_fp16 = const()[name = string("const_154_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_4828_cast_fp16 = mul(x = x2_51_cast_fp16, y = const_154_promoted_to_fp16)[name = string("op_4828_cast_fp16")];
            int32 var_4830 = const()[name = string("op_4830"), val = int32(-1)];
            bool var_4831_interleave_0 = const()[name = string("op_4831_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_4831_cast_fp16 = concat(axis = var_4830, interleave = var_4831_interleave_0, values = (var_4828_cast_fp16, x1_51_cast_fp16))[name = string("op_4831_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4834_cast_fp16 = mul(x = var_4831_cast_fp16, y = var_1105_cast_fp16)[name = string("op_4834_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_49_cast_fp16 = add(x = var_4807_cast_fp16, y = var_4834_cast_fp16)[name = string("x_49_cast_fp16")];
            tensor<int32, [4]> var_4841 = const()[name = string("op_4841"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_4853 = const()[name = string("op_4853"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_4842_cast_fp16 = transpose(perm = var_4841, x = x_49_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 1024, 1, 16]> var_4854_cast_fp16 = reshape(shape = var_4853, x = var_4842_cast_fp16)[name = string("op_4854_cast_fp16")];
            tensor<int32, [1]> var_4856_axes_0 = const()[name = string("op_4856_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4856_cast_fp16 = squeeze(axes = var_4856_axes_0, x = var_4854_cast_fp16)[name = string("op_4856_cast_fp16")];
            bool var_4857_transpose_x_0 = const()[name = string("op_4857_transpose_x_0"), val = bool(false)];
            bool var_4857_transpose_y_0 = const()[name = string("op_4857_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4857_cast_fp16 = matmul(transpose_x = var_4857_transpose_x_0, transpose_y = var_4857_transpose_y_0, x = var_4856_cast_fp16, y = block_cache_map)[name = string("op_4857_cast_fp16")];
            tensor<int32, [1]> nk_full_25_axes_0 = const()[name = string("nk_full_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_25_cast_fp16 = expand_dims(axes = nk_full_25_axes_0, x = var_4857_cast_fp16)[name = string("nk_full_25_cast_fp16")];
            tensor<int32, [4]> var_4876 = const()[name = string("op_4876"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_12_cast_fp16 = transpose(perm = transpose_12_perm_0, x = v_25_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 1024, 1, 16]> var_4877_cast_fp16 = reshape(shape = var_4876, x = transpose_12_cast_fp16)[name = string("op_4877_cast_fp16")];
            tensor<int32, [1]> var_4879_axes_0 = const()[name = string("op_4879_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_4879_cast_fp16 = squeeze(axes = var_4879_axes_0, x = var_4877_cast_fp16)[name = string("op_4879_cast_fp16")];
            bool var_4880_transpose_x_0 = const()[name = string("op_4880_transpose_x_0"), val = bool(false)];
            bool var_4880_transpose_y_0 = const()[name = string("op_4880_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_4880_cast_fp16 = matmul(transpose_x = var_4880_transpose_x_0, transpose_y = var_4880_transpose_y_0, x = var_4879_cast_fp16, y = block_cache_map)[name = string("op_4880_cast_fp16")];
            tensor<int32, [1]> nv_full_25_axes_0 = const()[name = string("nv_full_25_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_25_cast_fp16 = expand_dims(axes = nv_full_25_axes_0, x = var_4880_cast_fp16)[name = string("nv_full_25_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4886_cast_fp16 = mul(x = lkc_49_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4886_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4887_cast_fp16 = mul(x = nk_full_25_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4887_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_51_cast_fp16 = add(x = var_4886_cast_fp16, y = var_4887_cast_fp16)[name = string("lkc_51_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4893_cast_fp16 = mul(x = lvc_49_cast_fp16, y = var_1189_cast_fp16)[name = string("op_4893_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_4894_cast_fp16 = mul(x = nv_full_25_cast_fp16, y = update_mask_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_51_cast_fp16 = add(x = var_4893_cast_fp16, y = var_4894_cast_fp16)[name = string("lvc_51_cast_fp16")];
            tensor<int32, [1]> var_4898_axes_0 = const()[name = string("op_4898_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4898_cast_fp16 = squeeze(axes = var_4898_axes_0, x = lkc_51_cast_fp16)[name = string("op_4898_cast_fp16")];
            tensor<int32, [4]> var_4902 = const()[name = string("op_4902"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_25_cast_fp16 = reshape(shape = var_4902, x = var_4898_cast_fp16)[name = string("kc_25_cast_fp16")];
            tensor<int32, [1]> var_4905_axes_0 = const()[name = string("op_4905_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_4905_cast_fp16 = squeeze(axes = var_4905_axes_0, x = lvc_51_cast_fp16)[name = string("op_4905_cast_fp16")];
            tensor<int32, [4]> var_4909 = const()[name = string("op_4909"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_25_cast_fp16 = reshape(shape = var_4909, x = var_4905_cast_fp16)[name = string("vc_25_cast_fp16")];
            bool var_4911_transpose_x_0 = const()[name = string("op_4911_transpose_x_0"), val = bool(false)];
            bool var_4911_transpose_y_0 = const()[name = string("op_4911_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_4911_cast_fp16 = matmul(transpose_x = var_4911_transpose_x_0, transpose_y = var_4911_transpose_y_0, x = q_77_cast_fp16, y = kc_25_cast_fp16)[name = string("op_4911_cast_fp16")];
            fp16 _inversed_scores_73_y_0_to_fp16 = const()[name = string("_inversed_scores_73_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_73_cast_fp16 = mul(x = var_4911_cast_fp16, y = _inversed_scores_73_y_0_to_fp16)[name = string("_inversed_scores_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_75_cast_fp16 = add(x = _inversed_scores_73_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_75_cast_fp16")];
            int32 var_4940 = const()[name = string("op_4940"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_4942_cast_fp16 = softmax(axis = var_4940, x = scores_75_cast_fp16)[name = string("op_4942_cast_fp16")];
            bool attn_out_49_transpose_x_1 = const()[name = string("attn_out_49_transpose_x_1"), val = bool(false)];
            bool attn_out_49_transpose_y_1 = const()[name = string("attn_out_49_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_49_cast_fp16 = matmul(transpose_x = attn_out_49_transpose_x_1, transpose_y = attn_out_49_transpose_y_1, x = var_4942_cast_fp16, y = vc_25_cast_fp16)[name = string("attn_out_49_cast_fp16")];
            tensor<int32, [4]> var_4954_perm_0 = const()[name = string("op_4954_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4956 = const()[name = string("op_4956"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_4954_cast_fp16 = transpose(perm = var_4954_perm_0, x = attn_out_49_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 1024]> input_101_cast_fp16 = reshape(shape = var_4956, x = var_4954_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_12_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_12_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434193856)))];
            tensor<fp16, [1, 16, 1024]> linear_87_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_self_attn_o_proj_weight_to_fp16, x = input_101_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_199_cast_fp16 = add(x = hidden_states_191_cast_fp16, y = linear_87_cast_fp16)[name = string("hidden_states_199_cast_fp16")];
            fp16 var_4966_promoted_to_fp16 = const()[name = string("op_4966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4972_cast_fp16 = pow(x = hidden_states_199_cast_fp16, y = var_4966_promoted_to_fp16)[name = string("op_4972_cast_fp16")];
            tensor<int32, [1]> variance_51_axes_0 = const()[name = string("variance_51_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_51_keep_dims_0 = const()[name = string("variance_51_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_51_cast_fp16 = reduce_mean(axes = variance_51_axes_0, keep_dims = variance_51_keep_dims_0, x = var_4972_cast_fp16)[name = string("variance_51_cast_fp16")];
            fp16 var_4975_to_fp16 = const()[name = string("op_4975_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_4976_cast_fp16 = add(x = variance_51_cast_fp16, y = var_4975_to_fp16)[name = string("op_4976_cast_fp16")];
            fp32 var_4977_epsilon_0 = const()[name = string("op_4977_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4977_cast_fp16 = rsqrt(epsilon = var_4977_epsilon_0, x = var_4976_cast_fp16)[name = string("op_4977_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_203_cast_fp16 = mul(x = hidden_states_199_cast_fp16, y = var_4977_cast_fp16)[name = string("hidden_states_203_cast_fp16")];
            tensor<fp16, [1024]> layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436291072)))];
            tensor<fp16, [1, 16, 1024]> input_103_cast_fp16 = mul(x = layers_12_post_attention_layernorm_weight_to_fp16, y = hidden_states_203_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_12_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(436293184)))];
            tensor<fp16, [1, 16, 4096]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_gate_proj_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_4990_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_4990_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_12_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(444681856)))];
            tensor<fp16, [1, 16, 4096]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_12_mlp_up_proj_weight_to_fp16, x = input_103_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_107_cast_fp16 = mul(x = var_4990_cast_fp16, y = linear_89_cast_fp16)[name = string("input_107_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_12_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_12_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453070528)))];
            tensor<fp16, [1, 16, 1024]> linear_90_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_12_mlp_down_proj_weight_to_fp16, x = input_107_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_207_cast_fp16 = add(x = hidden_states_199_cast_fp16, y = linear_90_cast_fp16)[name = string("hidden_states_207_cast_fp16")];
            tensor<int32, [4]> lkc_53_begin_0 = const()[name = string("lkc_53_begin_0"), val = tensor<int32, [4]>([0, 13312, 0, 0])];
            tensor<int32, [4]> lkc_53_end_0 = const()[name = string("lkc_53_end_0"), val = tensor<int32, [4]>([1, 14336, 1, 1024])];
            tensor<bool, [4]> lkc_53_end_mask_0 = const()[name = string("lkc_53_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_53_cast_fp16 = slice_by_index(begin = lkc_53_begin_0, end = lkc_53_end_0, end_mask = lkc_53_end_mask_0, x = key_cache)[name = string("lkc_53_cast_fp16")];
            tensor<int32, [4]> lvc_53_begin_0 = const()[name = string("lvc_53_begin_0"), val = tensor<int32, [4]>([0, 13312, 0, 0])];
            tensor<int32, [4]> lvc_53_end_0 = const()[name = string("lvc_53_end_0"), val = tensor<int32, [4]>([1, 14336, 1, 1024])];
            tensor<bool, [4]> lvc_53_end_mask_0 = const()[name = string("lvc_53_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_53_cast_fp16 = slice_by_index(begin = lvc_53_begin_0, end = lvc_53_end_0, end_mask = lvc_53_end_mask_0, x = value_cache)[name = string("lvc_53_cast_fp16")];
            fp16 var_5021_promoted_to_fp16 = const()[name = string("op_5021_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5027_cast_fp16 = pow(x = hidden_states_207_cast_fp16, y = var_5021_promoted_to_fp16)[name = string("op_5027_cast_fp16")];
            tensor<int32, [1]> variance_53_axes_0 = const()[name = string("variance_53_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_53_keep_dims_0 = const()[name = string("variance_53_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_53_cast_fp16 = reduce_mean(axes = variance_53_axes_0, keep_dims = variance_53_keep_dims_0, x = var_5027_cast_fp16)[name = string("variance_53_cast_fp16")];
            fp16 var_5030_to_fp16 = const()[name = string("op_5030_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5031_cast_fp16 = add(x = variance_53_cast_fp16, y = var_5030_to_fp16)[name = string("op_5031_cast_fp16")];
            fp32 var_5032_epsilon_0 = const()[name = string("op_5032_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5032_cast_fp16 = rsqrt(epsilon = var_5032_epsilon_0, x = var_5031_cast_fp16)[name = string("op_5032_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_211_cast_fp16 = mul(x = hidden_states_207_cast_fp16, y = var_5032_cast_fp16)[name = string("hidden_states_211_cast_fp16")];
            tensor<fp16, [1024]> layers_13_input_layernorm_weight_to_fp16 = const()[name = string("layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461459200)))];
            tensor<fp16, [1, 16, 1024]> hidden_53_cast_fp16 = mul(x = layers_13_input_layernorm_weight_to_fp16, y = hidden_states_211_cast_fp16)[name = string("hidden_53_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(461461312)))];
            tensor<fp16, [1, 16, 1024]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_q_proj_weight_to_fp16, x = hidden_53_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5056 = const()[name = string("op_5056"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_79_cast_fp16 = reshape(shape = var_5056, x = linear_91_cast_fp16)[name = string("q_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(463558528)))];
            tensor<fp16, [1, 16, 1024]> linear_92_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_k_proj_weight_to_fp16, x = hidden_53_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5063 = const()[name = string("op_5063"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_53_cast_fp16 = reshape(shape = var_5063, x = linear_92_cast_fp16)[name = string("k_53_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465655744)))];
            tensor<fp16, [1, 16, 1024]> linear_93_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_v_proj_weight_to_fp16, x = hidden_53_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5070 = const()[name = string("op_5070"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_27_cast_fp16 = reshape(shape = var_5070, x = linear_93_cast_fp16)[name = string("v_27_cast_fp16")];
            tensor<int32, [4]> q_81_perm_0 = const()[name = string("q_81_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_55_perm_0 = const()[name = string("k_55_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_13_perm_0 = const()[name = string("transpose_13_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_81_cast_fp16 = transpose(perm = q_81_perm_0, x = q_79_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 64]> var_5083_cast_fp16 = mul(x = q_81_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5083_cast_fp16")];
            tensor<int32, [4]> x1_53_begin_0 = const()[name = string("x1_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_53_end_0 = const()[name = string("x1_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_53_end_mask_0 = const()[name = string("x1_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_53_cast_fp16 = slice_by_index(begin = x1_53_begin_0, end = x1_53_end_0, end_mask = x1_53_end_mask_0, x = q_81_cast_fp16)[name = string("x1_53_cast_fp16")];
            tensor<int32, [4]> x2_53_begin_0 = const()[name = string("x2_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_53_end_0 = const()[name = string("x2_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_53_end_mask_0 = const()[name = string("x2_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_53_cast_fp16 = slice_by_index(begin = x2_53_begin_0, end = x2_53_end_0, end_mask = x2_53_end_mask_0, x = q_81_cast_fp16)[name = string("x2_53_cast_fp16")];
            fp16 const_163_promoted_to_fp16 = const()[name = string("const_163_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5104_cast_fp16 = mul(x = x2_53_cast_fp16, y = const_163_promoted_to_fp16)[name = string("op_5104_cast_fp16")];
            int32 var_5106 = const()[name = string("op_5106"), val = int32(-1)];
            bool var_5107_interleave_0 = const()[name = string("op_5107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5107_cast_fp16 = concat(axis = var_5106, interleave = var_5107_interleave_0, values = (var_5104_cast_fp16, x1_53_cast_fp16))[name = string("op_5107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5110_cast_fp16 = mul(x = var_5107_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5110_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_83_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5110_cast_fp16)[name = string("q_83_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_55_cast_fp16 = transpose(perm = k_55_perm_0, x = k_53_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 16, 16, 64]> var_5115_cast_fp16 = mul(x = k_55_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5115_cast_fp16")];
            tensor<int32, [4]> x1_55_begin_0 = const()[name = string("x1_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_55_end_0 = const()[name = string("x1_55_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_55_end_mask_0 = const()[name = string("x1_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_55_cast_fp16 = slice_by_index(begin = x1_55_begin_0, end = x1_55_end_0, end_mask = x1_55_end_mask_0, x = k_55_cast_fp16)[name = string("x1_55_cast_fp16")];
            tensor<int32, [4]> x2_55_begin_0 = const()[name = string("x2_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_55_end_0 = const()[name = string("x2_55_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_55_end_mask_0 = const()[name = string("x2_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_55_cast_fp16 = slice_by_index(begin = x2_55_begin_0, end = x2_55_end_0, end_mask = x2_55_end_mask_0, x = k_55_cast_fp16)[name = string("x2_55_cast_fp16")];
            fp16 const_166_promoted_to_fp16 = const()[name = string("const_166_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5136_cast_fp16 = mul(x = x2_55_cast_fp16, y = const_166_promoted_to_fp16)[name = string("op_5136_cast_fp16")];
            int32 var_5138 = const()[name = string("op_5138"), val = int32(-1)];
            bool var_5139_interleave_0 = const()[name = string("op_5139_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5139_cast_fp16 = concat(axis = var_5138, interleave = var_5139_interleave_0, values = (var_5136_cast_fp16, x1_55_cast_fp16))[name = string("op_5139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5142_cast_fp16 = mul(x = var_5139_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5142_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_53_cast_fp16 = add(x = var_5115_cast_fp16, y = var_5142_cast_fp16)[name = string("x_53_cast_fp16")];
            tensor<int32, [4]> var_5149 = const()[name = string("op_5149"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5161 = const()[name = string("op_5161"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_5150_cast_fp16 = transpose(perm = var_5149, x = x_53_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 1024, 1, 16]> var_5162_cast_fp16 = reshape(shape = var_5161, x = var_5150_cast_fp16)[name = string("op_5162_cast_fp16")];
            tensor<int32, [1]> var_5164_axes_0 = const()[name = string("op_5164_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5164_cast_fp16 = squeeze(axes = var_5164_axes_0, x = var_5162_cast_fp16)[name = string("op_5164_cast_fp16")];
            bool var_5165_transpose_x_0 = const()[name = string("op_5165_transpose_x_0"), val = bool(false)];
            bool var_5165_transpose_y_0 = const()[name = string("op_5165_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5165_cast_fp16 = matmul(transpose_x = var_5165_transpose_x_0, transpose_y = var_5165_transpose_y_0, x = var_5164_cast_fp16, y = block_cache_map)[name = string("op_5165_cast_fp16")];
            tensor<int32, [1]> nk_full_27_axes_0 = const()[name = string("nk_full_27_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_27_cast_fp16 = expand_dims(axes = nk_full_27_axes_0, x = var_5165_cast_fp16)[name = string("nk_full_27_cast_fp16")];
            tensor<int32, [4]> var_5184 = const()[name = string("op_5184"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_13_cast_fp16 = transpose(perm = transpose_13_perm_0, x = v_27_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 1024, 1, 16]> var_5185_cast_fp16 = reshape(shape = var_5184, x = transpose_13_cast_fp16)[name = string("op_5185_cast_fp16")];
            tensor<int32, [1]> var_5187_axes_0 = const()[name = string("op_5187_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5187_cast_fp16 = squeeze(axes = var_5187_axes_0, x = var_5185_cast_fp16)[name = string("op_5187_cast_fp16")];
            bool var_5188_transpose_x_0 = const()[name = string("op_5188_transpose_x_0"), val = bool(false)];
            bool var_5188_transpose_y_0 = const()[name = string("op_5188_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5188_cast_fp16 = matmul(transpose_x = var_5188_transpose_x_0, transpose_y = var_5188_transpose_y_0, x = var_5187_cast_fp16, y = block_cache_map)[name = string("op_5188_cast_fp16")];
            tensor<int32, [1]> nv_full_27_axes_0 = const()[name = string("nv_full_27_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_27_cast_fp16 = expand_dims(axes = nv_full_27_axes_0, x = var_5188_cast_fp16)[name = string("nv_full_27_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5194_cast_fp16 = mul(x = lkc_53_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5194_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5195_cast_fp16 = mul(x = nk_full_27_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5195_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_55_cast_fp16 = add(x = var_5194_cast_fp16, y = var_5195_cast_fp16)[name = string("lkc_55_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5201_cast_fp16 = mul(x = lvc_53_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5201_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5202_cast_fp16 = mul(x = nv_full_27_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5202_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_55_cast_fp16 = add(x = var_5201_cast_fp16, y = var_5202_cast_fp16)[name = string("lvc_55_cast_fp16")];
            tensor<int32, [1]> var_5206_axes_0 = const()[name = string("op_5206_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5206_cast_fp16 = squeeze(axes = var_5206_axes_0, x = lkc_55_cast_fp16)[name = string("op_5206_cast_fp16")];
            tensor<int32, [4]> var_5210 = const()[name = string("op_5210"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_27_cast_fp16 = reshape(shape = var_5210, x = var_5206_cast_fp16)[name = string("kc_27_cast_fp16")];
            tensor<int32, [1]> var_5213_axes_0 = const()[name = string("op_5213_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5213_cast_fp16 = squeeze(axes = var_5213_axes_0, x = lvc_55_cast_fp16)[name = string("op_5213_cast_fp16")];
            tensor<int32, [4]> var_5217 = const()[name = string("op_5217"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_27_cast_fp16 = reshape(shape = var_5217, x = var_5213_cast_fp16)[name = string("vc_27_cast_fp16")];
            bool var_5219_transpose_x_0 = const()[name = string("op_5219_transpose_x_0"), val = bool(false)];
            bool var_5219_transpose_y_0 = const()[name = string("op_5219_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_5219_cast_fp16 = matmul(transpose_x = var_5219_transpose_x_0, transpose_y = var_5219_transpose_y_0, x = q_83_cast_fp16, y = kc_27_cast_fp16)[name = string("op_5219_cast_fp16")];
            fp16 _inversed_scores_79_y_0_to_fp16 = const()[name = string("_inversed_scores_79_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_79_cast_fp16 = mul(x = var_5219_cast_fp16, y = _inversed_scores_79_y_0_to_fp16)[name = string("_inversed_scores_79_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_81_cast_fp16 = add(x = _inversed_scores_79_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_81_cast_fp16")];
            int32 var_5248 = const()[name = string("op_5248"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_5250_cast_fp16 = softmax(axis = var_5248, x = scores_81_cast_fp16)[name = string("op_5250_cast_fp16")];
            bool attn_out_53_transpose_x_1 = const()[name = string("attn_out_53_transpose_x_1"), val = bool(false)];
            bool attn_out_53_transpose_y_1 = const()[name = string("attn_out_53_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_53_cast_fp16 = matmul(transpose_x = attn_out_53_transpose_x_1, transpose_y = attn_out_53_transpose_y_1, x = var_5250_cast_fp16, y = vc_27_cast_fp16)[name = string("attn_out_53_cast_fp16")];
            tensor<int32, [4]> var_5262_perm_0 = const()[name = string("op_5262_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5264 = const()[name = string("op_5264"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_5262_cast_fp16 = transpose(perm = var_5262_perm_0, x = attn_out_53_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 1024]> input_109_cast_fp16 = reshape(shape = var_5264, x = var_5262_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_13_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_13_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(467752960)))];
            tensor<fp16, [1, 16, 1024]> linear_94_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_self_attn_o_proj_weight_to_fp16, x = input_109_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_215_cast_fp16 = add(x = hidden_states_207_cast_fp16, y = linear_94_cast_fp16)[name = string("hidden_states_215_cast_fp16")];
            fp16 var_5274_promoted_to_fp16 = const()[name = string("op_5274_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5280_cast_fp16 = pow(x = hidden_states_215_cast_fp16, y = var_5274_promoted_to_fp16)[name = string("op_5280_cast_fp16")];
            tensor<int32, [1]> variance_55_axes_0 = const()[name = string("variance_55_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_55_keep_dims_0 = const()[name = string("variance_55_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_55_cast_fp16 = reduce_mean(axes = variance_55_axes_0, keep_dims = variance_55_keep_dims_0, x = var_5280_cast_fp16)[name = string("variance_55_cast_fp16")];
            fp16 var_5283_to_fp16 = const()[name = string("op_5283_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5284_cast_fp16 = add(x = variance_55_cast_fp16, y = var_5283_to_fp16)[name = string("op_5284_cast_fp16")];
            fp32 var_5285_epsilon_0 = const()[name = string("op_5285_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5285_cast_fp16 = rsqrt(epsilon = var_5285_epsilon_0, x = var_5284_cast_fp16)[name = string("op_5285_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_219_cast_fp16 = mul(x = hidden_states_215_cast_fp16, y = var_5285_cast_fp16)[name = string("hidden_states_219_cast_fp16")];
            tensor<fp16, [1024]> layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469850176)))];
            tensor<fp16, [1, 16, 1024]> input_111_cast_fp16 = mul(x = layers_13_post_attention_layernorm_weight_to_fp16, y = hidden_states_219_cast_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_13_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(469852288)))];
            tensor<fp16, [1, 16, 4096]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_gate_proj_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_5298_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_5298_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_13_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478240960)))];
            tensor<fp16, [1, 16, 4096]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_13_mlp_up_proj_weight_to_fp16, x = input_111_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_115_cast_fp16 = mul(x = var_5298_cast_fp16, y = linear_96_cast_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_13_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_13_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486629632)))];
            tensor<fp16, [1, 16, 1024]> linear_97_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_13_mlp_down_proj_weight_to_fp16, x = input_115_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_223_cast_fp16 = add(x = hidden_states_215_cast_fp16, y = linear_97_cast_fp16)[name = string("hidden_states_223_cast_fp16")];
            tensor<int32, [4]> lkc_57_begin_0 = const()[name = string("lkc_57_begin_0"), val = tensor<int32, [4]>([0, 14336, 0, 0])];
            tensor<int32, [4]> lkc_57_end_0 = const()[name = string("lkc_57_end_0"), val = tensor<int32, [4]>([1, 15360, 1, 1024])];
            tensor<bool, [4]> lkc_57_end_mask_0 = const()[name = string("lkc_57_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_57_cast_fp16 = slice_by_index(begin = lkc_57_begin_0, end = lkc_57_end_0, end_mask = lkc_57_end_mask_0, x = key_cache)[name = string("lkc_57_cast_fp16")];
            tensor<int32, [4]> lvc_57_begin_0 = const()[name = string("lvc_57_begin_0"), val = tensor<int32, [4]>([0, 14336, 0, 0])];
            tensor<int32, [4]> lvc_57_end_0 = const()[name = string("lvc_57_end_0"), val = tensor<int32, [4]>([1, 15360, 1, 1024])];
            tensor<bool, [4]> lvc_57_end_mask_0 = const()[name = string("lvc_57_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_57_cast_fp16 = slice_by_index(begin = lvc_57_begin_0, end = lvc_57_end_0, end_mask = lvc_57_end_mask_0, x = value_cache)[name = string("lvc_57_cast_fp16")];
            fp16 var_5329_promoted_to_fp16 = const()[name = string("op_5329_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5335_cast_fp16 = pow(x = hidden_states_223_cast_fp16, y = var_5329_promoted_to_fp16)[name = string("op_5335_cast_fp16")];
            tensor<int32, [1]> variance_57_axes_0 = const()[name = string("variance_57_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_57_keep_dims_0 = const()[name = string("variance_57_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_57_cast_fp16 = reduce_mean(axes = variance_57_axes_0, keep_dims = variance_57_keep_dims_0, x = var_5335_cast_fp16)[name = string("variance_57_cast_fp16")];
            fp16 var_5338_to_fp16 = const()[name = string("op_5338_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5339_cast_fp16 = add(x = variance_57_cast_fp16, y = var_5338_to_fp16)[name = string("op_5339_cast_fp16")];
            fp32 var_5340_epsilon_0 = const()[name = string("op_5340_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5340_cast_fp16 = rsqrt(epsilon = var_5340_epsilon_0, x = var_5339_cast_fp16)[name = string("op_5340_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_227_cast_fp16 = mul(x = hidden_states_223_cast_fp16, y = var_5340_cast_fp16)[name = string("hidden_states_227_cast_fp16")];
            tensor<fp16, [1024]> layers_14_input_layernorm_weight_to_fp16 = const()[name = string("layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495018304)))];
            tensor<fp16, [1, 16, 1024]> hidden_57_cast_fp16 = mul(x = layers_14_input_layernorm_weight_to_fp16, y = hidden_states_227_cast_fp16)[name = string("hidden_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(495020416)))];
            tensor<fp16, [1, 16, 1024]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_q_proj_weight_to_fp16, x = hidden_57_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_5364 = const()[name = string("op_5364"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_85_cast_fp16 = reshape(shape = var_5364, x = linear_98_cast_fp16)[name = string("q_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497117632)))];
            tensor<fp16, [1, 16, 1024]> linear_99_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_k_proj_weight_to_fp16, x = hidden_57_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_5371 = const()[name = string("op_5371"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_57_cast_fp16 = reshape(shape = var_5371, x = linear_99_cast_fp16)[name = string("k_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(499214848)))];
            tensor<fp16, [1, 16, 1024]> linear_100_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_v_proj_weight_to_fp16, x = hidden_57_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5378 = const()[name = string("op_5378"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_29_cast_fp16 = reshape(shape = var_5378, x = linear_100_cast_fp16)[name = string("v_29_cast_fp16")];
            tensor<int32, [4]> q_87_perm_0 = const()[name = string("q_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_59_perm_0 = const()[name = string("k_59_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_14_perm_0 = const()[name = string("transpose_14_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_87_cast_fp16 = transpose(perm = q_87_perm_0, x = q_85_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 64]> var_5391_cast_fp16 = mul(x = q_87_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5391_cast_fp16")];
            tensor<int32, [4]> x1_57_begin_0 = const()[name = string("x1_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_57_end_0 = const()[name = string("x1_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_57_end_mask_0 = const()[name = string("x1_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_57_cast_fp16 = slice_by_index(begin = x1_57_begin_0, end = x1_57_end_0, end_mask = x1_57_end_mask_0, x = q_87_cast_fp16)[name = string("x1_57_cast_fp16")];
            tensor<int32, [4]> x2_57_begin_0 = const()[name = string("x2_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_57_end_0 = const()[name = string("x2_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_57_end_mask_0 = const()[name = string("x2_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_57_cast_fp16 = slice_by_index(begin = x2_57_begin_0, end = x2_57_end_0, end_mask = x2_57_end_mask_0, x = q_87_cast_fp16)[name = string("x2_57_cast_fp16")];
            fp16 const_175_promoted_to_fp16 = const()[name = string("const_175_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5412_cast_fp16 = mul(x = x2_57_cast_fp16, y = const_175_promoted_to_fp16)[name = string("op_5412_cast_fp16")];
            int32 var_5414 = const()[name = string("op_5414"), val = int32(-1)];
            bool var_5415_interleave_0 = const()[name = string("op_5415_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5415_cast_fp16 = concat(axis = var_5414, interleave = var_5415_interleave_0, values = (var_5412_cast_fp16, x1_57_cast_fp16))[name = string("op_5415_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5418_cast_fp16 = mul(x = var_5415_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5418_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_89_cast_fp16 = add(x = var_5391_cast_fp16, y = var_5418_cast_fp16)[name = string("q_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_59_cast_fp16 = transpose(perm = k_59_perm_0, x = k_57_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 16, 16, 64]> var_5423_cast_fp16 = mul(x = k_59_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5423_cast_fp16")];
            tensor<int32, [4]> x1_59_begin_0 = const()[name = string("x1_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_59_end_0 = const()[name = string("x1_59_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_59_end_mask_0 = const()[name = string("x1_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_59_cast_fp16 = slice_by_index(begin = x1_59_begin_0, end = x1_59_end_0, end_mask = x1_59_end_mask_0, x = k_59_cast_fp16)[name = string("x1_59_cast_fp16")];
            tensor<int32, [4]> x2_59_begin_0 = const()[name = string("x2_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_59_end_0 = const()[name = string("x2_59_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_59_end_mask_0 = const()[name = string("x2_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_59_cast_fp16 = slice_by_index(begin = x2_59_begin_0, end = x2_59_end_0, end_mask = x2_59_end_mask_0, x = k_59_cast_fp16)[name = string("x2_59_cast_fp16")];
            fp16 const_178_promoted_to_fp16 = const()[name = string("const_178_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5444_cast_fp16 = mul(x = x2_59_cast_fp16, y = const_178_promoted_to_fp16)[name = string("op_5444_cast_fp16")];
            int32 var_5446 = const()[name = string("op_5446"), val = int32(-1)];
            bool var_5447_interleave_0 = const()[name = string("op_5447_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5447_cast_fp16 = concat(axis = var_5446, interleave = var_5447_interleave_0, values = (var_5444_cast_fp16, x1_59_cast_fp16))[name = string("op_5447_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5450_cast_fp16 = mul(x = var_5447_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5450_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_57_cast_fp16 = add(x = var_5423_cast_fp16, y = var_5450_cast_fp16)[name = string("x_57_cast_fp16")];
            tensor<int32, [4]> var_5457 = const()[name = string("op_5457"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5469 = const()[name = string("op_5469"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_5458_cast_fp16 = transpose(perm = var_5457, x = x_57_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 1024, 1, 16]> var_5470_cast_fp16 = reshape(shape = var_5469, x = var_5458_cast_fp16)[name = string("op_5470_cast_fp16")];
            tensor<int32, [1]> var_5472_axes_0 = const()[name = string("op_5472_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5472_cast_fp16 = squeeze(axes = var_5472_axes_0, x = var_5470_cast_fp16)[name = string("op_5472_cast_fp16")];
            bool var_5473_transpose_x_0 = const()[name = string("op_5473_transpose_x_0"), val = bool(false)];
            bool var_5473_transpose_y_0 = const()[name = string("op_5473_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5473_cast_fp16 = matmul(transpose_x = var_5473_transpose_x_0, transpose_y = var_5473_transpose_y_0, x = var_5472_cast_fp16, y = block_cache_map)[name = string("op_5473_cast_fp16")];
            tensor<int32, [1]> nk_full_29_axes_0 = const()[name = string("nk_full_29_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_29_cast_fp16 = expand_dims(axes = nk_full_29_axes_0, x = var_5473_cast_fp16)[name = string("nk_full_29_cast_fp16")];
            tensor<int32, [4]> var_5492 = const()[name = string("op_5492"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_14_cast_fp16 = transpose(perm = transpose_14_perm_0, x = v_29_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 1024, 1, 16]> var_5493_cast_fp16 = reshape(shape = var_5492, x = transpose_14_cast_fp16)[name = string("op_5493_cast_fp16")];
            tensor<int32, [1]> var_5495_axes_0 = const()[name = string("op_5495_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5495_cast_fp16 = squeeze(axes = var_5495_axes_0, x = var_5493_cast_fp16)[name = string("op_5495_cast_fp16")];
            bool var_5496_transpose_x_0 = const()[name = string("op_5496_transpose_x_0"), val = bool(false)];
            bool var_5496_transpose_y_0 = const()[name = string("op_5496_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5496_cast_fp16 = matmul(transpose_x = var_5496_transpose_x_0, transpose_y = var_5496_transpose_y_0, x = var_5495_cast_fp16, y = block_cache_map)[name = string("op_5496_cast_fp16")];
            tensor<int32, [1]> nv_full_29_axes_0 = const()[name = string("nv_full_29_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_29_cast_fp16 = expand_dims(axes = nv_full_29_axes_0, x = var_5496_cast_fp16)[name = string("nv_full_29_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5502_cast_fp16 = mul(x = lkc_57_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5502_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5503_cast_fp16 = mul(x = nk_full_29_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5503_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_59_cast_fp16 = add(x = var_5502_cast_fp16, y = var_5503_cast_fp16)[name = string("lkc_59_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5509_cast_fp16 = mul(x = lvc_57_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5509_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5510_cast_fp16 = mul(x = nv_full_29_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5510_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_59_cast_fp16 = add(x = var_5509_cast_fp16, y = var_5510_cast_fp16)[name = string("lvc_59_cast_fp16")];
            tensor<int32, [1]> var_5514_axes_0 = const()[name = string("op_5514_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5514_cast_fp16 = squeeze(axes = var_5514_axes_0, x = lkc_59_cast_fp16)[name = string("op_5514_cast_fp16")];
            tensor<int32, [4]> var_5518 = const()[name = string("op_5518"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_29_cast_fp16 = reshape(shape = var_5518, x = var_5514_cast_fp16)[name = string("kc_29_cast_fp16")];
            tensor<int32, [1]> var_5521_axes_0 = const()[name = string("op_5521_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5521_cast_fp16 = squeeze(axes = var_5521_axes_0, x = lvc_59_cast_fp16)[name = string("op_5521_cast_fp16")];
            tensor<int32, [4]> var_5525 = const()[name = string("op_5525"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_29_cast_fp16 = reshape(shape = var_5525, x = var_5521_cast_fp16)[name = string("vc_29_cast_fp16")];
            bool var_5527_transpose_x_0 = const()[name = string("op_5527_transpose_x_0"), val = bool(false)];
            bool var_5527_transpose_y_0 = const()[name = string("op_5527_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_5527_cast_fp16 = matmul(transpose_x = var_5527_transpose_x_0, transpose_y = var_5527_transpose_y_0, x = q_89_cast_fp16, y = kc_29_cast_fp16)[name = string("op_5527_cast_fp16")];
            fp16 _inversed_scores_85_y_0_to_fp16 = const()[name = string("_inversed_scores_85_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_85_cast_fp16 = mul(x = var_5527_cast_fp16, y = _inversed_scores_85_y_0_to_fp16)[name = string("_inversed_scores_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_87_cast_fp16 = add(x = _inversed_scores_85_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_87_cast_fp16")];
            int32 var_5556 = const()[name = string("op_5556"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_5558_cast_fp16 = softmax(axis = var_5556, x = scores_87_cast_fp16)[name = string("op_5558_cast_fp16")];
            bool attn_out_57_transpose_x_1 = const()[name = string("attn_out_57_transpose_x_1"), val = bool(false)];
            bool attn_out_57_transpose_y_1 = const()[name = string("attn_out_57_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_57_cast_fp16 = matmul(transpose_x = attn_out_57_transpose_x_1, transpose_y = attn_out_57_transpose_y_1, x = var_5558_cast_fp16, y = vc_29_cast_fp16)[name = string("attn_out_57_cast_fp16")];
            tensor<int32, [4]> var_5570_perm_0 = const()[name = string("op_5570_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5572 = const()[name = string("op_5572"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_5570_cast_fp16 = transpose(perm = var_5570_perm_0, x = attn_out_57_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 1024]> input_117_cast_fp16 = reshape(shape = var_5572, x = var_5570_cast_fp16)[name = string("input_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_14_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_14_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(501312064)))];
            tensor<fp16, [1, 16, 1024]> linear_101_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_self_attn_o_proj_weight_to_fp16, x = input_117_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_231_cast_fp16 = add(x = hidden_states_223_cast_fp16, y = linear_101_cast_fp16)[name = string("hidden_states_231_cast_fp16")];
            fp16 var_5582_promoted_to_fp16 = const()[name = string("op_5582_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5588_cast_fp16 = pow(x = hidden_states_231_cast_fp16, y = var_5582_promoted_to_fp16)[name = string("op_5588_cast_fp16")];
            tensor<int32, [1]> variance_59_axes_0 = const()[name = string("variance_59_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_59_keep_dims_0 = const()[name = string("variance_59_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_59_cast_fp16 = reduce_mean(axes = variance_59_axes_0, keep_dims = variance_59_keep_dims_0, x = var_5588_cast_fp16)[name = string("variance_59_cast_fp16")];
            fp16 var_5591_to_fp16 = const()[name = string("op_5591_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5592_cast_fp16 = add(x = variance_59_cast_fp16, y = var_5591_to_fp16)[name = string("op_5592_cast_fp16")];
            fp32 var_5593_epsilon_0 = const()[name = string("op_5593_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5593_cast_fp16 = rsqrt(epsilon = var_5593_epsilon_0, x = var_5592_cast_fp16)[name = string("op_5593_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_235_cast_fp16 = mul(x = hidden_states_231_cast_fp16, y = var_5593_cast_fp16)[name = string("hidden_states_235_cast_fp16")];
            tensor<fp16, [1024]> layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503409280)))];
            tensor<fp16, [1, 16, 1024]> input_119_cast_fp16 = mul(x = layers_14_post_attention_layernorm_weight_to_fp16, y = hidden_states_235_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_14_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_14_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503411392)))];
            tensor<fp16, [1, 16, 4096]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_14_mlp_gate_proj_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_5606_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_14_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_14_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(511800064)))];
            tensor<fp16, [1, 16, 4096]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_14_mlp_up_proj_weight_to_fp16, x = input_119_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_123_cast_fp16 = mul(x = var_5606_cast_fp16, y = linear_103_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_14_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_14_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520188736)))];
            tensor<fp16, [1, 16, 1024]> linear_104_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_14_mlp_down_proj_weight_to_fp16, x = input_123_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_239_cast_fp16 = add(x = hidden_states_231_cast_fp16, y = linear_104_cast_fp16)[name = string("hidden_states_239_cast_fp16")];
            tensor<int32, [4]> lkc_61_begin_0 = const()[name = string("lkc_61_begin_0"), val = tensor<int32, [4]>([0, 15360, 0, 0])];
            tensor<int32, [4]> lkc_61_end_0 = const()[name = string("lkc_61_end_0"), val = tensor<int32, [4]>([1, 16384, 1, 1024])];
            tensor<bool, [4]> lkc_61_end_mask_0 = const()[name = string("lkc_61_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_61_cast_fp16 = slice_by_index(begin = lkc_61_begin_0, end = lkc_61_end_0, end_mask = lkc_61_end_mask_0, x = key_cache)[name = string("lkc_61_cast_fp16")];
            tensor<int32, [4]> lvc_61_begin_0 = const()[name = string("lvc_61_begin_0"), val = tensor<int32, [4]>([0, 15360, 0, 0])];
            tensor<int32, [4]> lvc_61_end_0 = const()[name = string("lvc_61_end_0"), val = tensor<int32, [4]>([1, 16384, 1, 1024])];
            tensor<bool, [4]> lvc_61_end_mask_0 = const()[name = string("lvc_61_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_61_cast_fp16 = slice_by_index(begin = lvc_61_begin_0, end = lvc_61_end_0, end_mask = lvc_61_end_mask_0, x = value_cache)[name = string("lvc_61_cast_fp16")];
            fp16 var_5637_promoted_to_fp16 = const()[name = string("op_5637_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5643_cast_fp16 = pow(x = hidden_states_239_cast_fp16, y = var_5637_promoted_to_fp16)[name = string("op_5643_cast_fp16")];
            tensor<int32, [1]> variance_61_axes_0 = const()[name = string("variance_61_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_61_keep_dims_0 = const()[name = string("variance_61_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_61_cast_fp16 = reduce_mean(axes = variance_61_axes_0, keep_dims = variance_61_keep_dims_0, x = var_5643_cast_fp16)[name = string("variance_61_cast_fp16")];
            fp16 var_5646_to_fp16 = const()[name = string("op_5646_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5647_cast_fp16 = add(x = variance_61_cast_fp16, y = var_5646_to_fp16)[name = string("op_5647_cast_fp16")];
            fp32 var_5648_epsilon_0 = const()[name = string("op_5648_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5648_cast_fp16 = rsqrt(epsilon = var_5648_epsilon_0, x = var_5647_cast_fp16)[name = string("op_5648_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_243_cast_fp16 = mul(x = hidden_states_239_cast_fp16, y = var_5648_cast_fp16)[name = string("hidden_states_243_cast_fp16")];
            tensor<fp16, [1024]> layers_15_input_layernorm_weight_to_fp16 = const()[name = string("layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528577408)))];
            tensor<fp16, [1, 16, 1024]> hidden_61_cast_fp16 = mul(x = layers_15_input_layernorm_weight_to_fp16, y = hidden_states_243_cast_fp16)[name = string("hidden_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528579520)))];
            tensor<fp16, [1, 16, 1024]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_q_proj_weight_to_fp16, x = hidden_61_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_5672 = const()[name = string("op_5672"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_91_cast_fp16 = reshape(shape = var_5672, x = linear_105_cast_fp16)[name = string("q_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(530676736)))];
            tensor<fp16, [1, 16, 1024]> linear_106_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_k_proj_weight_to_fp16, x = hidden_61_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_5679 = const()[name = string("op_5679"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_61_cast_fp16 = reshape(shape = var_5679, x = linear_106_cast_fp16)[name = string("k_61_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(532773952)))];
            tensor<fp16, [1, 16, 1024]> linear_107_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_v_proj_weight_to_fp16, x = hidden_61_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_5686 = const()[name = string("op_5686"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_31_cast_fp16 = reshape(shape = var_5686, x = linear_107_cast_fp16)[name = string("v_31_cast_fp16")];
            tensor<int32, [4]> q_93_perm_0 = const()[name = string("q_93_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_63_perm_0 = const()[name = string("k_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_15_perm_0 = const()[name = string("transpose_15_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_93_cast_fp16 = transpose(perm = q_93_perm_0, x = q_91_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 64]> var_5699_cast_fp16 = mul(x = q_93_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5699_cast_fp16")];
            tensor<int32, [4]> x1_61_begin_0 = const()[name = string("x1_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_61_end_0 = const()[name = string("x1_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_61_end_mask_0 = const()[name = string("x1_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_61_cast_fp16 = slice_by_index(begin = x1_61_begin_0, end = x1_61_end_0, end_mask = x1_61_end_mask_0, x = q_93_cast_fp16)[name = string("x1_61_cast_fp16")];
            tensor<int32, [4]> x2_61_begin_0 = const()[name = string("x2_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_61_end_0 = const()[name = string("x2_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_61_end_mask_0 = const()[name = string("x2_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_61_cast_fp16 = slice_by_index(begin = x2_61_begin_0, end = x2_61_end_0, end_mask = x2_61_end_mask_0, x = q_93_cast_fp16)[name = string("x2_61_cast_fp16")];
            fp16 const_187_promoted_to_fp16 = const()[name = string("const_187_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5720_cast_fp16 = mul(x = x2_61_cast_fp16, y = const_187_promoted_to_fp16)[name = string("op_5720_cast_fp16")];
            int32 var_5722 = const()[name = string("op_5722"), val = int32(-1)];
            bool var_5723_interleave_0 = const()[name = string("op_5723_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5723_cast_fp16 = concat(axis = var_5722, interleave = var_5723_interleave_0, values = (var_5720_cast_fp16, x1_61_cast_fp16))[name = string("op_5723_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5726_cast_fp16 = mul(x = var_5723_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5726_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_95_cast_fp16 = add(x = var_5699_cast_fp16, y = var_5726_cast_fp16)[name = string("q_95_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_63_cast_fp16 = transpose(perm = k_63_perm_0, x = k_61_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 16, 16, 64]> var_5731_cast_fp16 = mul(x = k_63_cast_fp16, y = var_1078_cast_fp16)[name = string("op_5731_cast_fp16")];
            tensor<int32, [4]> x1_63_begin_0 = const()[name = string("x1_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_63_end_0 = const()[name = string("x1_63_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_63_end_mask_0 = const()[name = string("x1_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_63_cast_fp16 = slice_by_index(begin = x1_63_begin_0, end = x1_63_end_0, end_mask = x1_63_end_mask_0, x = k_63_cast_fp16)[name = string("x1_63_cast_fp16")];
            tensor<int32, [4]> x2_63_begin_0 = const()[name = string("x2_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_63_end_0 = const()[name = string("x2_63_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_63_end_mask_0 = const()[name = string("x2_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_63_cast_fp16 = slice_by_index(begin = x2_63_begin_0, end = x2_63_end_0, end_mask = x2_63_end_mask_0, x = k_63_cast_fp16)[name = string("x2_63_cast_fp16")];
            fp16 const_190_promoted_to_fp16 = const()[name = string("const_190_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_5752_cast_fp16 = mul(x = x2_63_cast_fp16, y = const_190_promoted_to_fp16)[name = string("op_5752_cast_fp16")];
            int32 var_5754 = const()[name = string("op_5754"), val = int32(-1)];
            bool var_5755_interleave_0 = const()[name = string("op_5755_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_5755_cast_fp16 = concat(axis = var_5754, interleave = var_5755_interleave_0, values = (var_5752_cast_fp16, x1_63_cast_fp16))[name = string("op_5755_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5758_cast_fp16 = mul(x = var_5755_cast_fp16, y = var_1105_cast_fp16)[name = string("op_5758_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_61_cast_fp16 = add(x = var_5731_cast_fp16, y = var_5758_cast_fp16)[name = string("x_61_cast_fp16")];
            tensor<int32, [4]> var_5765 = const()[name = string("op_5765"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_5777 = const()[name = string("op_5777"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_5766_cast_fp16 = transpose(perm = var_5765, x = x_61_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 1024, 1, 16]> var_5778_cast_fp16 = reshape(shape = var_5777, x = var_5766_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<int32, [1]> var_5780_axes_0 = const()[name = string("op_5780_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5780_cast_fp16 = squeeze(axes = var_5780_axes_0, x = var_5778_cast_fp16)[name = string("op_5780_cast_fp16")];
            bool var_5781_transpose_x_0 = const()[name = string("op_5781_transpose_x_0"), val = bool(false)];
            bool var_5781_transpose_y_0 = const()[name = string("op_5781_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5781_cast_fp16 = matmul(transpose_x = var_5781_transpose_x_0, transpose_y = var_5781_transpose_y_0, x = var_5780_cast_fp16, y = block_cache_map)[name = string("op_5781_cast_fp16")];
            tensor<int32, [1]> nk_full_31_axes_0 = const()[name = string("nk_full_31_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_31_cast_fp16 = expand_dims(axes = nk_full_31_axes_0, x = var_5781_cast_fp16)[name = string("nk_full_31_cast_fp16")];
            tensor<int32, [4]> var_5800 = const()[name = string("op_5800"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_15_cast_fp16 = transpose(perm = transpose_15_perm_0, x = v_31_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 1024, 1, 16]> var_5801_cast_fp16 = reshape(shape = var_5800, x = transpose_15_cast_fp16)[name = string("op_5801_cast_fp16")];
            tensor<int32, [1]> var_5803_axes_0 = const()[name = string("op_5803_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_5803_cast_fp16 = squeeze(axes = var_5803_axes_0, x = var_5801_cast_fp16)[name = string("op_5803_cast_fp16")];
            bool var_5804_transpose_x_0 = const()[name = string("op_5804_transpose_x_0"), val = bool(false)];
            bool var_5804_transpose_y_0 = const()[name = string("op_5804_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_5804_cast_fp16 = matmul(transpose_x = var_5804_transpose_x_0, transpose_y = var_5804_transpose_y_0, x = var_5803_cast_fp16, y = block_cache_map)[name = string("op_5804_cast_fp16")];
            tensor<int32, [1]> nv_full_31_axes_0 = const()[name = string("nv_full_31_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_31_cast_fp16 = expand_dims(axes = nv_full_31_axes_0, x = var_5804_cast_fp16)[name = string("nv_full_31_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5810_cast_fp16 = mul(x = lkc_61_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5810_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5811_cast_fp16 = mul(x = nk_full_31_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5811_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_63_cast_fp16 = add(x = var_5810_cast_fp16, y = var_5811_cast_fp16)[name = string("lkc_63_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5817_cast_fp16 = mul(x = lvc_61_cast_fp16, y = var_1189_cast_fp16)[name = string("op_5817_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_5818_cast_fp16 = mul(x = nv_full_31_cast_fp16, y = update_mask_cast_fp16)[name = string("op_5818_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_63_cast_fp16 = add(x = var_5817_cast_fp16, y = var_5818_cast_fp16)[name = string("lvc_63_cast_fp16")];
            tensor<int32, [1]> var_5822_axes_0 = const()[name = string("op_5822_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5822_cast_fp16 = squeeze(axes = var_5822_axes_0, x = lkc_63_cast_fp16)[name = string("op_5822_cast_fp16")];
            tensor<int32, [4]> var_5826 = const()[name = string("op_5826"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_31_cast_fp16 = reshape(shape = var_5826, x = var_5822_cast_fp16)[name = string("kc_31_cast_fp16")];
            tensor<int32, [1]> var_5829_axes_0 = const()[name = string("op_5829_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_5829_cast_fp16 = squeeze(axes = var_5829_axes_0, x = lvc_63_cast_fp16)[name = string("op_5829_cast_fp16")];
            tensor<int32, [4]> var_5833 = const()[name = string("op_5833"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_31_cast_fp16 = reshape(shape = var_5833, x = var_5829_cast_fp16)[name = string("vc_31_cast_fp16")];
            bool var_5835_transpose_x_0 = const()[name = string("op_5835_transpose_x_0"), val = bool(false)];
            bool var_5835_transpose_y_0 = const()[name = string("op_5835_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_5835_cast_fp16 = matmul(transpose_x = var_5835_transpose_x_0, transpose_y = var_5835_transpose_y_0, x = q_95_cast_fp16, y = kc_31_cast_fp16)[name = string("op_5835_cast_fp16")];
            fp16 _inversed_scores_91_y_0_to_fp16 = const()[name = string("_inversed_scores_91_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_91_cast_fp16 = mul(x = var_5835_cast_fp16, y = _inversed_scores_91_y_0_to_fp16)[name = string("_inversed_scores_91_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_93_cast_fp16 = add(x = _inversed_scores_91_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_93_cast_fp16")];
            int32 var_5864 = const()[name = string("op_5864"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_5866_cast_fp16 = softmax(axis = var_5864, x = scores_93_cast_fp16)[name = string("op_5866_cast_fp16")];
            bool attn_out_61_transpose_x_1 = const()[name = string("attn_out_61_transpose_x_1"), val = bool(false)];
            bool attn_out_61_transpose_y_1 = const()[name = string("attn_out_61_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_61_cast_fp16 = matmul(transpose_x = attn_out_61_transpose_x_1, transpose_y = attn_out_61_transpose_y_1, x = var_5866_cast_fp16, y = vc_31_cast_fp16)[name = string("attn_out_61_cast_fp16")];
            tensor<int32, [4]> var_5878_perm_0 = const()[name = string("op_5878_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5880 = const()[name = string("op_5880"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_5878_cast_fp16 = transpose(perm = var_5878_perm_0, x = attn_out_61_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 1024]> input_125_cast_fp16 = reshape(shape = var_5880, x = var_5878_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_15_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_15_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(534871168)))];
            tensor<fp16, [1, 16, 1024]> linear_108_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_self_attn_o_proj_weight_to_fp16, x = input_125_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_247_cast_fp16 = add(x = hidden_states_239_cast_fp16, y = linear_108_cast_fp16)[name = string("hidden_states_247_cast_fp16")];
            fp16 var_5890_promoted_to_fp16 = const()[name = string("op_5890_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5896_cast_fp16 = pow(x = hidden_states_247_cast_fp16, y = var_5890_promoted_to_fp16)[name = string("op_5896_cast_fp16")];
            tensor<int32, [1]> variance_63_axes_0 = const()[name = string("variance_63_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_63_keep_dims_0 = const()[name = string("variance_63_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_63_cast_fp16 = reduce_mean(axes = variance_63_axes_0, keep_dims = variance_63_keep_dims_0, x = var_5896_cast_fp16)[name = string("variance_63_cast_fp16")];
            fp16 var_5899_to_fp16 = const()[name = string("op_5899_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5900_cast_fp16 = add(x = variance_63_cast_fp16, y = var_5899_to_fp16)[name = string("op_5900_cast_fp16")];
            fp32 var_5901_epsilon_0 = const()[name = string("op_5901_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5901_cast_fp16 = rsqrt(epsilon = var_5901_epsilon_0, x = var_5900_cast_fp16)[name = string("op_5901_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_251_cast_fp16 = mul(x = hidden_states_247_cast_fp16, y = var_5901_cast_fp16)[name = string("hidden_states_251_cast_fp16")];
            tensor<fp16, [1024]> layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536968384)))];
            tensor<fp16, [1, 16, 1024]> input_127_cast_fp16 = mul(x = layers_15_post_attention_layernorm_weight_to_fp16, y = hidden_states_251_cast_fp16)[name = string("input_127_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_15_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_15_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536970496)))];
            tensor<fp16, [1, 16, 4096]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_15_mlp_gate_proj_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_5914_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_5914_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_15_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_15_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(545359168)))];
            tensor<fp16, [1, 16, 4096]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_15_mlp_up_proj_weight_to_fp16, x = input_127_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_131_cast_fp16 = mul(x = var_5914_cast_fp16, y = linear_110_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_15_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_15_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553747840)))];
            tensor<fp16, [1, 16, 1024]> linear_111_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_15_mlp_down_proj_weight_to_fp16, x = input_131_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_255_cast_fp16 = add(x = hidden_states_247_cast_fp16, y = linear_111_cast_fp16)[name = string("hidden_states_255_cast_fp16")];
            tensor<int32, [4]> lkc_65_begin_0 = const()[name = string("lkc_65_begin_0"), val = tensor<int32, [4]>([0, 16384, 0, 0])];
            tensor<int32, [4]> lkc_65_end_0 = const()[name = string("lkc_65_end_0"), val = tensor<int32, [4]>([1, 17408, 1, 1024])];
            tensor<bool, [4]> lkc_65_end_mask_0 = const()[name = string("lkc_65_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_65_cast_fp16 = slice_by_index(begin = lkc_65_begin_0, end = lkc_65_end_0, end_mask = lkc_65_end_mask_0, x = key_cache)[name = string("lkc_65_cast_fp16")];
            tensor<int32, [4]> lvc_65_begin_0 = const()[name = string("lvc_65_begin_0"), val = tensor<int32, [4]>([0, 16384, 0, 0])];
            tensor<int32, [4]> lvc_65_end_0 = const()[name = string("lvc_65_end_0"), val = tensor<int32, [4]>([1, 17408, 1, 1024])];
            tensor<bool, [4]> lvc_65_end_mask_0 = const()[name = string("lvc_65_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_65_cast_fp16 = slice_by_index(begin = lvc_65_begin_0, end = lvc_65_end_0, end_mask = lvc_65_end_mask_0, x = value_cache)[name = string("lvc_65_cast_fp16")];
            fp16 var_5945_promoted_to_fp16 = const()[name = string("op_5945_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5951_cast_fp16 = pow(x = hidden_states_255_cast_fp16, y = var_5945_promoted_to_fp16)[name = string("op_5951_cast_fp16")];
            tensor<int32, [1]> variance_65_axes_0 = const()[name = string("variance_65_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_65_keep_dims_0 = const()[name = string("variance_65_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_65_cast_fp16 = reduce_mean(axes = variance_65_axes_0, keep_dims = variance_65_keep_dims_0, x = var_5951_cast_fp16)[name = string("variance_65_cast_fp16")];
            fp16 var_5954_to_fp16 = const()[name = string("op_5954_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_5955_cast_fp16 = add(x = variance_65_cast_fp16, y = var_5954_to_fp16)[name = string("op_5955_cast_fp16")];
            fp32 var_5956_epsilon_0 = const()[name = string("op_5956_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5956_cast_fp16 = rsqrt(epsilon = var_5956_epsilon_0, x = var_5955_cast_fp16)[name = string("op_5956_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_259_cast_fp16 = mul(x = hidden_states_255_cast_fp16, y = var_5956_cast_fp16)[name = string("hidden_states_259_cast_fp16")];
            tensor<fp16, [1024]> layers_16_input_layernorm_weight_to_fp16 = const()[name = string("layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562136512)))];
            tensor<fp16, [1, 16, 1024]> hidden_65_cast_fp16 = mul(x = layers_16_input_layernorm_weight_to_fp16, y = hidden_states_259_cast_fp16)[name = string("hidden_65_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(562138624)))];
            tensor<fp16, [1, 16, 1024]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_q_proj_weight_to_fp16, x = hidden_65_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_5980 = const()[name = string("op_5980"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_97_cast_fp16 = reshape(shape = var_5980, x = linear_112_cast_fp16)[name = string("q_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(564235840)))];
            tensor<fp16, [1, 16, 1024]> linear_113_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_k_proj_weight_to_fp16, x = hidden_65_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_5987 = const()[name = string("op_5987"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_65_cast_fp16 = reshape(shape = var_5987, x = linear_113_cast_fp16)[name = string("k_65_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566333056)))];
            tensor<fp16, [1, 16, 1024]> linear_114_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_v_proj_weight_to_fp16, x = hidden_65_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_5994 = const()[name = string("op_5994"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_33_cast_fp16 = reshape(shape = var_5994, x = linear_114_cast_fp16)[name = string("v_33_cast_fp16")];
            tensor<int32, [4]> q_99_perm_0 = const()[name = string("q_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_67_perm_0 = const()[name = string("k_67_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_16_perm_0 = const()[name = string("transpose_16_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_99_cast_fp16 = transpose(perm = q_99_perm_0, x = q_97_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 64]> var_6007_cast_fp16 = mul(x = q_99_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6007_cast_fp16")];
            tensor<int32, [4]> x1_65_begin_0 = const()[name = string("x1_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_65_end_0 = const()[name = string("x1_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_65_end_mask_0 = const()[name = string("x1_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_65_cast_fp16 = slice_by_index(begin = x1_65_begin_0, end = x1_65_end_0, end_mask = x1_65_end_mask_0, x = q_99_cast_fp16)[name = string("x1_65_cast_fp16")];
            tensor<int32, [4]> x2_65_begin_0 = const()[name = string("x2_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_65_end_0 = const()[name = string("x2_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_65_end_mask_0 = const()[name = string("x2_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_65_cast_fp16 = slice_by_index(begin = x2_65_begin_0, end = x2_65_end_0, end_mask = x2_65_end_mask_0, x = q_99_cast_fp16)[name = string("x2_65_cast_fp16")];
            fp16 const_199_promoted_to_fp16 = const()[name = string("const_199_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6028_cast_fp16 = mul(x = x2_65_cast_fp16, y = const_199_promoted_to_fp16)[name = string("op_6028_cast_fp16")];
            int32 var_6030 = const()[name = string("op_6030"), val = int32(-1)];
            bool var_6031_interleave_0 = const()[name = string("op_6031_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6031_cast_fp16 = concat(axis = var_6030, interleave = var_6031_interleave_0, values = (var_6028_cast_fp16, x1_65_cast_fp16))[name = string("op_6031_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6034_cast_fp16 = mul(x = var_6031_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_101_cast_fp16 = add(x = var_6007_cast_fp16, y = var_6034_cast_fp16)[name = string("q_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_67_cast_fp16 = transpose(perm = k_67_perm_0, x = k_65_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 16, 16, 64]> var_6039_cast_fp16 = mul(x = k_67_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6039_cast_fp16")];
            tensor<int32, [4]> x1_67_begin_0 = const()[name = string("x1_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_67_end_0 = const()[name = string("x1_67_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_67_end_mask_0 = const()[name = string("x1_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_67_cast_fp16 = slice_by_index(begin = x1_67_begin_0, end = x1_67_end_0, end_mask = x1_67_end_mask_0, x = k_67_cast_fp16)[name = string("x1_67_cast_fp16")];
            tensor<int32, [4]> x2_67_begin_0 = const()[name = string("x2_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_67_end_0 = const()[name = string("x2_67_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_67_end_mask_0 = const()[name = string("x2_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_67_cast_fp16 = slice_by_index(begin = x2_67_begin_0, end = x2_67_end_0, end_mask = x2_67_end_mask_0, x = k_67_cast_fp16)[name = string("x2_67_cast_fp16")];
            fp16 const_202_promoted_to_fp16 = const()[name = string("const_202_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6060_cast_fp16 = mul(x = x2_67_cast_fp16, y = const_202_promoted_to_fp16)[name = string("op_6060_cast_fp16")];
            int32 var_6062 = const()[name = string("op_6062"), val = int32(-1)];
            bool var_6063_interleave_0 = const()[name = string("op_6063_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6063_cast_fp16 = concat(axis = var_6062, interleave = var_6063_interleave_0, values = (var_6060_cast_fp16, x1_67_cast_fp16))[name = string("op_6063_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6066_cast_fp16 = mul(x = var_6063_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6066_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_65_cast_fp16 = add(x = var_6039_cast_fp16, y = var_6066_cast_fp16)[name = string("x_65_cast_fp16")];
            tensor<int32, [4]> var_6073 = const()[name = string("op_6073"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6085 = const()[name = string("op_6085"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_6074_cast_fp16 = transpose(perm = var_6073, x = x_65_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 1024, 1, 16]> var_6086_cast_fp16 = reshape(shape = var_6085, x = var_6074_cast_fp16)[name = string("op_6086_cast_fp16")];
            tensor<int32, [1]> var_6088_axes_0 = const()[name = string("op_6088_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6088_cast_fp16 = squeeze(axes = var_6088_axes_0, x = var_6086_cast_fp16)[name = string("op_6088_cast_fp16")];
            bool var_6089_transpose_x_0 = const()[name = string("op_6089_transpose_x_0"), val = bool(false)];
            bool var_6089_transpose_y_0 = const()[name = string("op_6089_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6089_cast_fp16 = matmul(transpose_x = var_6089_transpose_x_0, transpose_y = var_6089_transpose_y_0, x = var_6088_cast_fp16, y = block_cache_map)[name = string("op_6089_cast_fp16")];
            tensor<int32, [1]> nk_full_33_axes_0 = const()[name = string("nk_full_33_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_33_cast_fp16 = expand_dims(axes = nk_full_33_axes_0, x = var_6089_cast_fp16)[name = string("nk_full_33_cast_fp16")];
            tensor<int32, [4]> var_6108 = const()[name = string("op_6108"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_16_cast_fp16 = transpose(perm = transpose_16_perm_0, x = v_33_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 1024, 1, 16]> var_6109_cast_fp16 = reshape(shape = var_6108, x = transpose_16_cast_fp16)[name = string("op_6109_cast_fp16")];
            tensor<int32, [1]> var_6111_axes_0 = const()[name = string("op_6111_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6111_cast_fp16 = squeeze(axes = var_6111_axes_0, x = var_6109_cast_fp16)[name = string("op_6111_cast_fp16")];
            bool var_6112_transpose_x_0 = const()[name = string("op_6112_transpose_x_0"), val = bool(false)];
            bool var_6112_transpose_y_0 = const()[name = string("op_6112_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6112_cast_fp16 = matmul(transpose_x = var_6112_transpose_x_0, transpose_y = var_6112_transpose_y_0, x = var_6111_cast_fp16, y = block_cache_map)[name = string("op_6112_cast_fp16")];
            tensor<int32, [1]> nv_full_33_axes_0 = const()[name = string("nv_full_33_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_33_cast_fp16 = expand_dims(axes = nv_full_33_axes_0, x = var_6112_cast_fp16)[name = string("nv_full_33_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6118_cast_fp16 = mul(x = lkc_65_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6118_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6119_cast_fp16 = mul(x = nk_full_33_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6119_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_67_cast_fp16 = add(x = var_6118_cast_fp16, y = var_6119_cast_fp16)[name = string("lkc_67_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6125_cast_fp16 = mul(x = lvc_65_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6125_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6126_cast_fp16 = mul(x = nv_full_33_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_67_cast_fp16 = add(x = var_6125_cast_fp16, y = var_6126_cast_fp16)[name = string("lvc_67_cast_fp16")];
            tensor<int32, [1]> var_6130_axes_0 = const()[name = string("op_6130_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6130_cast_fp16 = squeeze(axes = var_6130_axes_0, x = lkc_67_cast_fp16)[name = string("op_6130_cast_fp16")];
            tensor<int32, [4]> var_6134 = const()[name = string("op_6134"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_33_cast_fp16 = reshape(shape = var_6134, x = var_6130_cast_fp16)[name = string("kc_33_cast_fp16")];
            tensor<int32, [1]> var_6137_axes_0 = const()[name = string("op_6137_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6137_cast_fp16 = squeeze(axes = var_6137_axes_0, x = lvc_67_cast_fp16)[name = string("op_6137_cast_fp16")];
            tensor<int32, [4]> var_6141 = const()[name = string("op_6141"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_33_cast_fp16 = reshape(shape = var_6141, x = var_6137_cast_fp16)[name = string("vc_33_cast_fp16")];
            bool var_6143_transpose_x_0 = const()[name = string("op_6143_transpose_x_0"), val = bool(false)];
            bool var_6143_transpose_y_0 = const()[name = string("op_6143_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_6143_cast_fp16 = matmul(transpose_x = var_6143_transpose_x_0, transpose_y = var_6143_transpose_y_0, x = q_101_cast_fp16, y = kc_33_cast_fp16)[name = string("op_6143_cast_fp16")];
            fp16 _inversed_scores_97_y_0_to_fp16 = const()[name = string("_inversed_scores_97_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_97_cast_fp16 = mul(x = var_6143_cast_fp16, y = _inversed_scores_97_y_0_to_fp16)[name = string("_inversed_scores_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_99_cast_fp16 = add(x = _inversed_scores_97_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_99_cast_fp16")];
            int32 var_6172 = const()[name = string("op_6172"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_6174_cast_fp16 = softmax(axis = var_6172, x = scores_99_cast_fp16)[name = string("op_6174_cast_fp16")];
            bool attn_out_65_transpose_x_1 = const()[name = string("attn_out_65_transpose_x_1"), val = bool(false)];
            bool attn_out_65_transpose_y_1 = const()[name = string("attn_out_65_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_65_cast_fp16 = matmul(transpose_x = attn_out_65_transpose_x_1, transpose_y = attn_out_65_transpose_y_1, x = var_6174_cast_fp16, y = vc_33_cast_fp16)[name = string("attn_out_65_cast_fp16")];
            tensor<int32, [4]> var_6186_perm_0 = const()[name = string("op_6186_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6188 = const()[name = string("op_6188"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_6186_cast_fp16 = transpose(perm = var_6186_perm_0, x = attn_out_65_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 1024]> input_133_cast_fp16 = reshape(shape = var_6188, x = var_6186_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_16_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_16_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(568430272)))];
            tensor<fp16, [1, 16, 1024]> linear_115_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_self_attn_o_proj_weight_to_fp16, x = input_133_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_263_cast_fp16 = add(x = hidden_states_255_cast_fp16, y = linear_115_cast_fp16)[name = string("hidden_states_263_cast_fp16")];
            fp16 var_6198_promoted_to_fp16 = const()[name = string("op_6198_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6204_cast_fp16 = pow(x = hidden_states_263_cast_fp16, y = var_6198_promoted_to_fp16)[name = string("op_6204_cast_fp16")];
            tensor<int32, [1]> variance_67_axes_0 = const()[name = string("variance_67_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_67_keep_dims_0 = const()[name = string("variance_67_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_67_cast_fp16 = reduce_mean(axes = variance_67_axes_0, keep_dims = variance_67_keep_dims_0, x = var_6204_cast_fp16)[name = string("variance_67_cast_fp16")];
            fp16 var_6207_to_fp16 = const()[name = string("op_6207_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6208_cast_fp16 = add(x = variance_67_cast_fp16, y = var_6207_to_fp16)[name = string("op_6208_cast_fp16")];
            fp32 var_6209_epsilon_0 = const()[name = string("op_6209_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6209_cast_fp16 = rsqrt(epsilon = var_6209_epsilon_0, x = var_6208_cast_fp16)[name = string("op_6209_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_267_cast_fp16 = mul(x = hidden_states_263_cast_fp16, y = var_6209_cast_fp16)[name = string("hidden_states_267_cast_fp16")];
            tensor<fp16, [1024]> layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570527488)))];
            tensor<fp16, [1, 16, 1024]> input_135_cast_fp16 = mul(x = layers_16_post_attention_layernorm_weight_to_fp16, y = hidden_states_267_cast_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_16_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_16_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(570529600)))];
            tensor<fp16, [1, 16, 4096]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_16_mlp_gate_proj_weight_to_fp16, x = input_135_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_6222_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_6222_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_16_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_16_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(578918272)))];
            tensor<fp16, [1, 16, 4096]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_16_mlp_up_proj_weight_to_fp16, x = input_135_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_139_cast_fp16 = mul(x = var_6222_cast_fp16, y = linear_117_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_16_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_16_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(587306944)))];
            tensor<fp16, [1, 16, 1024]> linear_118_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_16_mlp_down_proj_weight_to_fp16, x = input_139_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_271_cast_fp16 = add(x = hidden_states_263_cast_fp16, y = linear_118_cast_fp16)[name = string("hidden_states_271_cast_fp16")];
            tensor<int32, [4]> lkc_69_begin_0 = const()[name = string("lkc_69_begin_0"), val = tensor<int32, [4]>([0, 17408, 0, 0])];
            tensor<int32, [4]> lkc_69_end_0 = const()[name = string("lkc_69_end_0"), val = tensor<int32, [4]>([1, 18432, 1, 1024])];
            tensor<bool, [4]> lkc_69_end_mask_0 = const()[name = string("lkc_69_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_69_cast_fp16 = slice_by_index(begin = lkc_69_begin_0, end = lkc_69_end_0, end_mask = lkc_69_end_mask_0, x = key_cache)[name = string("lkc_69_cast_fp16")];
            tensor<int32, [4]> lvc_69_begin_0 = const()[name = string("lvc_69_begin_0"), val = tensor<int32, [4]>([0, 17408, 0, 0])];
            tensor<int32, [4]> lvc_69_end_0 = const()[name = string("lvc_69_end_0"), val = tensor<int32, [4]>([1, 18432, 1, 1024])];
            tensor<bool, [4]> lvc_69_end_mask_0 = const()[name = string("lvc_69_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_69_cast_fp16 = slice_by_index(begin = lvc_69_begin_0, end = lvc_69_end_0, end_mask = lvc_69_end_mask_0, x = value_cache)[name = string("lvc_69_cast_fp16")];
            fp16 var_6253_promoted_to_fp16 = const()[name = string("op_6253_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6259_cast_fp16 = pow(x = hidden_states_271_cast_fp16, y = var_6253_promoted_to_fp16)[name = string("op_6259_cast_fp16")];
            tensor<int32, [1]> variance_69_axes_0 = const()[name = string("variance_69_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_69_keep_dims_0 = const()[name = string("variance_69_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_69_cast_fp16 = reduce_mean(axes = variance_69_axes_0, keep_dims = variance_69_keep_dims_0, x = var_6259_cast_fp16)[name = string("variance_69_cast_fp16")];
            fp16 var_6262_to_fp16 = const()[name = string("op_6262_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6263_cast_fp16 = add(x = variance_69_cast_fp16, y = var_6262_to_fp16)[name = string("op_6263_cast_fp16")];
            fp32 var_6264_epsilon_0 = const()[name = string("op_6264_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6264_cast_fp16 = rsqrt(epsilon = var_6264_epsilon_0, x = var_6263_cast_fp16)[name = string("op_6264_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_275_cast_fp16 = mul(x = hidden_states_271_cast_fp16, y = var_6264_cast_fp16)[name = string("hidden_states_275_cast_fp16")];
            tensor<fp16, [1024]> layers_17_input_layernorm_weight_to_fp16 = const()[name = string("layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595695616)))];
            tensor<fp16, [1, 16, 1024]> hidden_69_cast_fp16 = mul(x = layers_17_input_layernorm_weight_to_fp16, y = hidden_states_275_cast_fp16)[name = string("hidden_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(595697728)))];
            tensor<fp16, [1, 16, 1024]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_q_proj_weight_to_fp16, x = hidden_69_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6288 = const()[name = string("op_6288"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_103_cast_fp16 = reshape(shape = var_6288, x = linear_119_cast_fp16)[name = string("q_103_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597794944)))];
            tensor<fp16, [1, 16, 1024]> linear_120_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_k_proj_weight_to_fp16, x = hidden_69_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6295 = const()[name = string("op_6295"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_69_cast_fp16 = reshape(shape = var_6295, x = linear_120_cast_fp16)[name = string("k_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599892160)))];
            tensor<fp16, [1, 16, 1024]> linear_121_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_v_proj_weight_to_fp16, x = hidden_69_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_6302 = const()[name = string("op_6302"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_35_cast_fp16 = reshape(shape = var_6302, x = linear_121_cast_fp16)[name = string("v_35_cast_fp16")];
            tensor<int32, [4]> q_105_perm_0 = const()[name = string("q_105_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_71_perm_0 = const()[name = string("k_71_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_17_perm_0 = const()[name = string("transpose_17_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_105_cast_fp16 = transpose(perm = q_105_perm_0, x = q_103_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 64]> var_6315_cast_fp16 = mul(x = q_105_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6315_cast_fp16")];
            tensor<int32, [4]> x1_69_begin_0 = const()[name = string("x1_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_69_end_0 = const()[name = string("x1_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_69_end_mask_0 = const()[name = string("x1_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_69_cast_fp16 = slice_by_index(begin = x1_69_begin_0, end = x1_69_end_0, end_mask = x1_69_end_mask_0, x = q_105_cast_fp16)[name = string("x1_69_cast_fp16")];
            tensor<int32, [4]> x2_69_begin_0 = const()[name = string("x2_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_69_end_0 = const()[name = string("x2_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_69_end_mask_0 = const()[name = string("x2_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_69_cast_fp16 = slice_by_index(begin = x2_69_begin_0, end = x2_69_end_0, end_mask = x2_69_end_mask_0, x = q_105_cast_fp16)[name = string("x2_69_cast_fp16")];
            fp16 const_211_promoted_to_fp16 = const()[name = string("const_211_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6336_cast_fp16 = mul(x = x2_69_cast_fp16, y = const_211_promoted_to_fp16)[name = string("op_6336_cast_fp16")];
            int32 var_6338 = const()[name = string("op_6338"), val = int32(-1)];
            bool var_6339_interleave_0 = const()[name = string("op_6339_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6339_cast_fp16 = concat(axis = var_6338, interleave = var_6339_interleave_0, values = (var_6336_cast_fp16, x1_69_cast_fp16))[name = string("op_6339_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6342_cast_fp16 = mul(x = var_6339_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6342_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_107_cast_fp16 = add(x = var_6315_cast_fp16, y = var_6342_cast_fp16)[name = string("q_107_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_71_cast_fp16 = transpose(perm = k_71_perm_0, x = k_69_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 16, 16, 64]> var_6347_cast_fp16 = mul(x = k_71_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6347_cast_fp16")];
            tensor<int32, [4]> x1_71_begin_0 = const()[name = string("x1_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_71_end_0 = const()[name = string("x1_71_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_71_end_mask_0 = const()[name = string("x1_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_71_cast_fp16 = slice_by_index(begin = x1_71_begin_0, end = x1_71_end_0, end_mask = x1_71_end_mask_0, x = k_71_cast_fp16)[name = string("x1_71_cast_fp16")];
            tensor<int32, [4]> x2_71_begin_0 = const()[name = string("x2_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_71_end_0 = const()[name = string("x2_71_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_71_end_mask_0 = const()[name = string("x2_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_71_cast_fp16 = slice_by_index(begin = x2_71_begin_0, end = x2_71_end_0, end_mask = x2_71_end_mask_0, x = k_71_cast_fp16)[name = string("x2_71_cast_fp16")];
            fp16 const_214_promoted_to_fp16 = const()[name = string("const_214_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6368_cast_fp16 = mul(x = x2_71_cast_fp16, y = const_214_promoted_to_fp16)[name = string("op_6368_cast_fp16")];
            int32 var_6370 = const()[name = string("op_6370"), val = int32(-1)];
            bool var_6371_interleave_0 = const()[name = string("op_6371_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6371_cast_fp16 = concat(axis = var_6370, interleave = var_6371_interleave_0, values = (var_6368_cast_fp16, x1_71_cast_fp16))[name = string("op_6371_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6374_cast_fp16 = mul(x = var_6371_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6374_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_69_cast_fp16 = add(x = var_6347_cast_fp16, y = var_6374_cast_fp16)[name = string("x_69_cast_fp16")];
            tensor<int32, [4]> var_6381 = const()[name = string("op_6381"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6393 = const()[name = string("op_6393"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_6382_cast_fp16 = transpose(perm = var_6381, x = x_69_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 1024, 1, 16]> var_6394_cast_fp16 = reshape(shape = var_6393, x = var_6382_cast_fp16)[name = string("op_6394_cast_fp16")];
            tensor<int32, [1]> var_6396_axes_0 = const()[name = string("op_6396_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6396_cast_fp16 = squeeze(axes = var_6396_axes_0, x = var_6394_cast_fp16)[name = string("op_6396_cast_fp16")];
            bool var_6397_transpose_x_0 = const()[name = string("op_6397_transpose_x_0"), val = bool(false)];
            bool var_6397_transpose_y_0 = const()[name = string("op_6397_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6397_cast_fp16 = matmul(transpose_x = var_6397_transpose_x_0, transpose_y = var_6397_transpose_y_0, x = var_6396_cast_fp16, y = block_cache_map)[name = string("op_6397_cast_fp16")];
            tensor<int32, [1]> nk_full_35_axes_0 = const()[name = string("nk_full_35_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_35_cast_fp16 = expand_dims(axes = nk_full_35_axes_0, x = var_6397_cast_fp16)[name = string("nk_full_35_cast_fp16")];
            tensor<int32, [4]> var_6416 = const()[name = string("op_6416"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_17_cast_fp16 = transpose(perm = transpose_17_perm_0, x = v_35_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 1024, 1, 16]> var_6417_cast_fp16 = reshape(shape = var_6416, x = transpose_17_cast_fp16)[name = string("op_6417_cast_fp16")];
            tensor<int32, [1]> var_6419_axes_0 = const()[name = string("op_6419_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6419_cast_fp16 = squeeze(axes = var_6419_axes_0, x = var_6417_cast_fp16)[name = string("op_6419_cast_fp16")];
            bool var_6420_transpose_x_0 = const()[name = string("op_6420_transpose_x_0"), val = bool(false)];
            bool var_6420_transpose_y_0 = const()[name = string("op_6420_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6420_cast_fp16 = matmul(transpose_x = var_6420_transpose_x_0, transpose_y = var_6420_transpose_y_0, x = var_6419_cast_fp16, y = block_cache_map)[name = string("op_6420_cast_fp16")];
            tensor<int32, [1]> nv_full_35_axes_0 = const()[name = string("nv_full_35_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_35_cast_fp16 = expand_dims(axes = nv_full_35_axes_0, x = var_6420_cast_fp16)[name = string("nv_full_35_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6426_cast_fp16 = mul(x = lkc_69_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6426_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6427_cast_fp16 = mul(x = nk_full_35_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6427_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_71_cast_fp16 = add(x = var_6426_cast_fp16, y = var_6427_cast_fp16)[name = string("lkc_71_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6433_cast_fp16 = mul(x = lvc_69_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6433_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6434_cast_fp16 = mul(x = nv_full_35_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_71_cast_fp16 = add(x = var_6433_cast_fp16, y = var_6434_cast_fp16)[name = string("lvc_71_cast_fp16")];
            tensor<int32, [1]> var_6438_axes_0 = const()[name = string("op_6438_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6438_cast_fp16 = squeeze(axes = var_6438_axes_0, x = lkc_71_cast_fp16)[name = string("op_6438_cast_fp16")];
            tensor<int32, [4]> var_6442 = const()[name = string("op_6442"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_35_cast_fp16 = reshape(shape = var_6442, x = var_6438_cast_fp16)[name = string("kc_35_cast_fp16")];
            tensor<int32, [1]> var_6445_axes_0 = const()[name = string("op_6445_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6445_cast_fp16 = squeeze(axes = var_6445_axes_0, x = lvc_71_cast_fp16)[name = string("op_6445_cast_fp16")];
            tensor<int32, [4]> var_6449 = const()[name = string("op_6449"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_35_cast_fp16 = reshape(shape = var_6449, x = var_6445_cast_fp16)[name = string("vc_35_cast_fp16")];
            bool var_6451_transpose_x_0 = const()[name = string("op_6451_transpose_x_0"), val = bool(false)];
            bool var_6451_transpose_y_0 = const()[name = string("op_6451_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_6451_cast_fp16 = matmul(transpose_x = var_6451_transpose_x_0, transpose_y = var_6451_transpose_y_0, x = q_107_cast_fp16, y = kc_35_cast_fp16)[name = string("op_6451_cast_fp16")];
            fp16 _inversed_scores_103_y_0_to_fp16 = const()[name = string("_inversed_scores_103_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_103_cast_fp16 = mul(x = var_6451_cast_fp16, y = _inversed_scores_103_y_0_to_fp16)[name = string("_inversed_scores_103_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_105_cast_fp16 = add(x = _inversed_scores_103_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_105_cast_fp16")];
            int32 var_6480 = const()[name = string("op_6480"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_6482_cast_fp16 = softmax(axis = var_6480, x = scores_105_cast_fp16)[name = string("op_6482_cast_fp16")];
            bool attn_out_69_transpose_x_1 = const()[name = string("attn_out_69_transpose_x_1"), val = bool(false)];
            bool attn_out_69_transpose_y_1 = const()[name = string("attn_out_69_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_69_cast_fp16 = matmul(transpose_x = attn_out_69_transpose_x_1, transpose_y = attn_out_69_transpose_y_1, x = var_6482_cast_fp16, y = vc_35_cast_fp16)[name = string("attn_out_69_cast_fp16")];
            tensor<int32, [4]> var_6494_perm_0 = const()[name = string("op_6494_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6496 = const()[name = string("op_6496"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_6494_cast_fp16 = transpose(perm = var_6494_perm_0, x = attn_out_69_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 1024]> input_141_cast_fp16 = reshape(shape = var_6496, x = var_6494_cast_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_17_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_17_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601989376)))];
            tensor<fp16, [1, 16, 1024]> linear_122_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_self_attn_o_proj_weight_to_fp16, x = input_141_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_279_cast_fp16 = add(x = hidden_states_271_cast_fp16, y = linear_122_cast_fp16)[name = string("hidden_states_279_cast_fp16")];
            fp16 var_6506_promoted_to_fp16 = const()[name = string("op_6506_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6512_cast_fp16 = pow(x = hidden_states_279_cast_fp16, y = var_6506_promoted_to_fp16)[name = string("op_6512_cast_fp16")];
            tensor<int32, [1]> variance_71_axes_0 = const()[name = string("variance_71_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_71_keep_dims_0 = const()[name = string("variance_71_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_71_cast_fp16 = reduce_mean(axes = variance_71_axes_0, keep_dims = variance_71_keep_dims_0, x = var_6512_cast_fp16)[name = string("variance_71_cast_fp16")];
            fp16 var_6515_to_fp16 = const()[name = string("op_6515_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6516_cast_fp16 = add(x = variance_71_cast_fp16, y = var_6515_to_fp16)[name = string("op_6516_cast_fp16")];
            fp32 var_6517_epsilon_0 = const()[name = string("op_6517_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6517_cast_fp16 = rsqrt(epsilon = var_6517_epsilon_0, x = var_6516_cast_fp16)[name = string("op_6517_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_283_cast_fp16 = mul(x = hidden_states_279_cast_fp16, y = var_6517_cast_fp16)[name = string("hidden_states_283_cast_fp16")];
            tensor<fp16, [1024]> layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604086592)))];
            tensor<fp16, [1, 16, 1024]> input_143_cast_fp16 = mul(x = layers_17_post_attention_layernorm_weight_to_fp16, y = hidden_states_283_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_17_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_17_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604088704)))];
            tensor<fp16, [1, 16, 4096]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_17_mlp_gate_proj_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_6530_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_6530_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_17_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_17_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612477376)))];
            tensor<fp16, [1, 16, 4096]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_17_mlp_up_proj_weight_to_fp16, x = input_143_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_147_cast_fp16 = mul(x = var_6530_cast_fp16, y = linear_124_cast_fp16)[name = string("input_147_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_17_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_17_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620866048)))];
            tensor<fp16, [1, 16, 1024]> linear_125_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_17_mlp_down_proj_weight_to_fp16, x = input_147_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_287_cast_fp16 = add(x = hidden_states_279_cast_fp16, y = linear_125_cast_fp16)[name = string("hidden_states_287_cast_fp16")];
            tensor<int32, [4]> lkc_73_begin_0 = const()[name = string("lkc_73_begin_0"), val = tensor<int32, [4]>([0, 18432, 0, 0])];
            tensor<int32, [4]> lkc_73_end_0 = const()[name = string("lkc_73_end_0"), val = tensor<int32, [4]>([1, 19456, 1, 1024])];
            tensor<bool, [4]> lkc_73_end_mask_0 = const()[name = string("lkc_73_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_73_cast_fp16 = slice_by_index(begin = lkc_73_begin_0, end = lkc_73_end_0, end_mask = lkc_73_end_mask_0, x = key_cache)[name = string("lkc_73_cast_fp16")];
            tensor<int32, [4]> lvc_73_begin_0 = const()[name = string("lvc_73_begin_0"), val = tensor<int32, [4]>([0, 18432, 0, 0])];
            tensor<int32, [4]> lvc_73_end_0 = const()[name = string("lvc_73_end_0"), val = tensor<int32, [4]>([1, 19456, 1, 1024])];
            tensor<bool, [4]> lvc_73_end_mask_0 = const()[name = string("lvc_73_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_73_cast_fp16 = slice_by_index(begin = lvc_73_begin_0, end = lvc_73_end_0, end_mask = lvc_73_end_mask_0, x = value_cache)[name = string("lvc_73_cast_fp16")];
            fp16 var_6561_promoted_to_fp16 = const()[name = string("op_6561_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6567_cast_fp16 = pow(x = hidden_states_287_cast_fp16, y = var_6561_promoted_to_fp16)[name = string("op_6567_cast_fp16")];
            tensor<int32, [1]> variance_73_axes_0 = const()[name = string("variance_73_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_73_keep_dims_0 = const()[name = string("variance_73_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_73_cast_fp16 = reduce_mean(axes = variance_73_axes_0, keep_dims = variance_73_keep_dims_0, x = var_6567_cast_fp16)[name = string("variance_73_cast_fp16")];
            fp16 var_6570_to_fp16 = const()[name = string("op_6570_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6571_cast_fp16 = add(x = variance_73_cast_fp16, y = var_6570_to_fp16)[name = string("op_6571_cast_fp16")];
            fp32 var_6572_epsilon_0 = const()[name = string("op_6572_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6572_cast_fp16 = rsqrt(epsilon = var_6572_epsilon_0, x = var_6571_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_291_cast_fp16 = mul(x = hidden_states_287_cast_fp16, y = var_6572_cast_fp16)[name = string("hidden_states_291_cast_fp16")];
            tensor<fp16, [1024]> layers_18_input_layernorm_weight_to_fp16 = const()[name = string("layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629254720)))];
            tensor<fp16, [1, 16, 1024]> hidden_73_cast_fp16 = mul(x = layers_18_input_layernorm_weight_to_fp16, y = hidden_states_291_cast_fp16)[name = string("hidden_73_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629256832)))];
            tensor<fp16, [1, 16, 1024]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_q_proj_weight_to_fp16, x = hidden_73_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_6596 = const()[name = string("op_6596"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_109_cast_fp16 = reshape(shape = var_6596, x = linear_126_cast_fp16)[name = string("q_109_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(631354048)))];
            tensor<fp16, [1, 16, 1024]> linear_127_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_k_proj_weight_to_fp16, x = hidden_73_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6603 = const()[name = string("op_6603"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_73_cast_fp16 = reshape(shape = var_6603, x = linear_127_cast_fp16)[name = string("k_73_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(633451264)))];
            tensor<fp16, [1, 16, 1024]> linear_128_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_v_proj_weight_to_fp16, x = hidden_73_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6610 = const()[name = string("op_6610"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_37_cast_fp16 = reshape(shape = var_6610, x = linear_128_cast_fp16)[name = string("v_37_cast_fp16")];
            tensor<int32, [4]> q_111_perm_0 = const()[name = string("q_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_75_perm_0 = const()[name = string("k_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_18_perm_0 = const()[name = string("transpose_18_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_111_cast_fp16 = transpose(perm = q_111_perm_0, x = q_109_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 64]> var_6623_cast_fp16 = mul(x = q_111_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6623_cast_fp16")];
            tensor<int32, [4]> x1_73_begin_0 = const()[name = string("x1_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_73_end_0 = const()[name = string("x1_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_73_end_mask_0 = const()[name = string("x1_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_73_cast_fp16 = slice_by_index(begin = x1_73_begin_0, end = x1_73_end_0, end_mask = x1_73_end_mask_0, x = q_111_cast_fp16)[name = string("x1_73_cast_fp16")];
            tensor<int32, [4]> x2_73_begin_0 = const()[name = string("x2_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_73_end_0 = const()[name = string("x2_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_73_end_mask_0 = const()[name = string("x2_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_73_cast_fp16 = slice_by_index(begin = x2_73_begin_0, end = x2_73_end_0, end_mask = x2_73_end_mask_0, x = q_111_cast_fp16)[name = string("x2_73_cast_fp16")];
            fp16 const_223_promoted_to_fp16 = const()[name = string("const_223_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6644_cast_fp16 = mul(x = x2_73_cast_fp16, y = const_223_promoted_to_fp16)[name = string("op_6644_cast_fp16")];
            int32 var_6646 = const()[name = string("op_6646"), val = int32(-1)];
            bool var_6647_interleave_0 = const()[name = string("op_6647_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6647_cast_fp16 = concat(axis = var_6646, interleave = var_6647_interleave_0, values = (var_6644_cast_fp16, x1_73_cast_fp16))[name = string("op_6647_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6650_cast_fp16 = mul(x = var_6647_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6650_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_113_cast_fp16 = add(x = var_6623_cast_fp16, y = var_6650_cast_fp16)[name = string("q_113_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_75_cast_fp16 = transpose(perm = k_75_perm_0, x = k_73_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 16, 16, 64]> var_6655_cast_fp16 = mul(x = k_75_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6655_cast_fp16")];
            tensor<int32, [4]> x1_75_begin_0 = const()[name = string("x1_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_75_end_0 = const()[name = string("x1_75_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_75_end_mask_0 = const()[name = string("x1_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_75_cast_fp16 = slice_by_index(begin = x1_75_begin_0, end = x1_75_end_0, end_mask = x1_75_end_mask_0, x = k_75_cast_fp16)[name = string("x1_75_cast_fp16")];
            tensor<int32, [4]> x2_75_begin_0 = const()[name = string("x2_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_75_end_0 = const()[name = string("x2_75_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_75_end_mask_0 = const()[name = string("x2_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_75_cast_fp16 = slice_by_index(begin = x2_75_begin_0, end = x2_75_end_0, end_mask = x2_75_end_mask_0, x = k_75_cast_fp16)[name = string("x2_75_cast_fp16")];
            fp16 const_226_promoted_to_fp16 = const()[name = string("const_226_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6676_cast_fp16 = mul(x = x2_75_cast_fp16, y = const_226_promoted_to_fp16)[name = string("op_6676_cast_fp16")];
            int32 var_6678 = const()[name = string("op_6678"), val = int32(-1)];
            bool var_6679_interleave_0 = const()[name = string("op_6679_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6679_cast_fp16 = concat(axis = var_6678, interleave = var_6679_interleave_0, values = (var_6676_cast_fp16, x1_75_cast_fp16))[name = string("op_6679_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6682_cast_fp16 = mul(x = var_6679_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6682_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_73_cast_fp16 = add(x = var_6655_cast_fp16, y = var_6682_cast_fp16)[name = string("x_73_cast_fp16")];
            tensor<int32, [4]> var_6689 = const()[name = string("op_6689"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_6701 = const()[name = string("op_6701"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_6690_cast_fp16 = transpose(perm = var_6689, x = x_73_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 1024, 1, 16]> var_6702_cast_fp16 = reshape(shape = var_6701, x = var_6690_cast_fp16)[name = string("op_6702_cast_fp16")];
            tensor<int32, [1]> var_6704_axes_0 = const()[name = string("op_6704_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6704_cast_fp16 = squeeze(axes = var_6704_axes_0, x = var_6702_cast_fp16)[name = string("op_6704_cast_fp16")];
            bool var_6705_transpose_x_0 = const()[name = string("op_6705_transpose_x_0"), val = bool(false)];
            bool var_6705_transpose_y_0 = const()[name = string("op_6705_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6705_cast_fp16 = matmul(transpose_x = var_6705_transpose_x_0, transpose_y = var_6705_transpose_y_0, x = var_6704_cast_fp16, y = block_cache_map)[name = string("op_6705_cast_fp16")];
            tensor<int32, [1]> nk_full_37_axes_0 = const()[name = string("nk_full_37_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_37_cast_fp16 = expand_dims(axes = nk_full_37_axes_0, x = var_6705_cast_fp16)[name = string("nk_full_37_cast_fp16")];
            tensor<int32, [4]> var_6724 = const()[name = string("op_6724"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_18_cast_fp16 = transpose(perm = transpose_18_perm_0, x = v_37_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 1024, 1, 16]> var_6725_cast_fp16 = reshape(shape = var_6724, x = transpose_18_cast_fp16)[name = string("op_6725_cast_fp16")];
            tensor<int32, [1]> var_6727_axes_0 = const()[name = string("op_6727_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_6727_cast_fp16 = squeeze(axes = var_6727_axes_0, x = var_6725_cast_fp16)[name = string("op_6727_cast_fp16")];
            bool var_6728_transpose_x_0 = const()[name = string("op_6728_transpose_x_0"), val = bool(false)];
            bool var_6728_transpose_y_0 = const()[name = string("op_6728_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_6728_cast_fp16 = matmul(transpose_x = var_6728_transpose_x_0, transpose_y = var_6728_transpose_y_0, x = var_6727_cast_fp16, y = block_cache_map)[name = string("op_6728_cast_fp16")];
            tensor<int32, [1]> nv_full_37_axes_0 = const()[name = string("nv_full_37_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_37_cast_fp16 = expand_dims(axes = nv_full_37_axes_0, x = var_6728_cast_fp16)[name = string("nv_full_37_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6734_cast_fp16 = mul(x = lkc_73_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6734_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6735_cast_fp16 = mul(x = nk_full_37_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6735_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_75_cast_fp16 = add(x = var_6734_cast_fp16, y = var_6735_cast_fp16)[name = string("lkc_75_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6741_cast_fp16 = mul(x = lvc_73_cast_fp16, y = var_1189_cast_fp16)[name = string("op_6741_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_6742_cast_fp16 = mul(x = nv_full_37_cast_fp16, y = update_mask_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_75_cast_fp16 = add(x = var_6741_cast_fp16, y = var_6742_cast_fp16)[name = string("lvc_75_cast_fp16")];
            tensor<int32, [1]> var_6746_axes_0 = const()[name = string("op_6746_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6746_cast_fp16 = squeeze(axes = var_6746_axes_0, x = lkc_75_cast_fp16)[name = string("op_6746_cast_fp16")];
            tensor<int32, [4]> var_6750 = const()[name = string("op_6750"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_37_cast_fp16 = reshape(shape = var_6750, x = var_6746_cast_fp16)[name = string("kc_37_cast_fp16")];
            tensor<int32, [1]> var_6753_axes_0 = const()[name = string("op_6753_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_6753_cast_fp16 = squeeze(axes = var_6753_axes_0, x = lvc_75_cast_fp16)[name = string("op_6753_cast_fp16")];
            tensor<int32, [4]> var_6757 = const()[name = string("op_6757"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_37_cast_fp16 = reshape(shape = var_6757, x = var_6753_cast_fp16)[name = string("vc_37_cast_fp16")];
            bool var_6759_transpose_x_0 = const()[name = string("op_6759_transpose_x_0"), val = bool(false)];
            bool var_6759_transpose_y_0 = const()[name = string("op_6759_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_6759_cast_fp16 = matmul(transpose_x = var_6759_transpose_x_0, transpose_y = var_6759_transpose_y_0, x = q_113_cast_fp16, y = kc_37_cast_fp16)[name = string("op_6759_cast_fp16")];
            fp16 _inversed_scores_109_y_0_to_fp16 = const()[name = string("_inversed_scores_109_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_109_cast_fp16 = mul(x = var_6759_cast_fp16, y = _inversed_scores_109_y_0_to_fp16)[name = string("_inversed_scores_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_111_cast_fp16 = add(x = _inversed_scores_109_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_111_cast_fp16")];
            int32 var_6788 = const()[name = string("op_6788"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_6790_cast_fp16 = softmax(axis = var_6788, x = scores_111_cast_fp16)[name = string("op_6790_cast_fp16")];
            bool attn_out_73_transpose_x_1 = const()[name = string("attn_out_73_transpose_x_1"), val = bool(false)];
            bool attn_out_73_transpose_y_1 = const()[name = string("attn_out_73_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_73_cast_fp16 = matmul(transpose_x = attn_out_73_transpose_x_1, transpose_y = attn_out_73_transpose_y_1, x = var_6790_cast_fp16, y = vc_37_cast_fp16)[name = string("attn_out_73_cast_fp16")];
            tensor<int32, [4]> var_6802_perm_0 = const()[name = string("op_6802_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6804 = const()[name = string("op_6804"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_6802_cast_fp16 = transpose(perm = var_6802_perm_0, x = attn_out_73_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 1024]> input_149_cast_fp16 = reshape(shape = var_6804, x = var_6802_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_18_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_18_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635548480)))];
            tensor<fp16, [1, 16, 1024]> linear_129_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_self_attn_o_proj_weight_to_fp16, x = input_149_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_295_cast_fp16 = add(x = hidden_states_287_cast_fp16, y = linear_129_cast_fp16)[name = string("hidden_states_295_cast_fp16")];
            fp16 var_6814_promoted_to_fp16 = const()[name = string("op_6814_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6820_cast_fp16 = pow(x = hidden_states_295_cast_fp16, y = var_6814_promoted_to_fp16)[name = string("op_6820_cast_fp16")];
            tensor<int32, [1]> variance_75_axes_0 = const()[name = string("variance_75_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_75_keep_dims_0 = const()[name = string("variance_75_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_75_cast_fp16 = reduce_mean(axes = variance_75_axes_0, keep_dims = variance_75_keep_dims_0, x = var_6820_cast_fp16)[name = string("variance_75_cast_fp16")];
            fp16 var_6823_to_fp16 = const()[name = string("op_6823_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6824_cast_fp16 = add(x = variance_75_cast_fp16, y = var_6823_to_fp16)[name = string("op_6824_cast_fp16")];
            fp32 var_6825_epsilon_0 = const()[name = string("op_6825_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6825_cast_fp16 = rsqrt(epsilon = var_6825_epsilon_0, x = var_6824_cast_fp16)[name = string("op_6825_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_299_cast_fp16 = mul(x = hidden_states_295_cast_fp16, y = var_6825_cast_fp16)[name = string("hidden_states_299_cast_fp16")];
            tensor<fp16, [1024]> layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637645696)))];
            tensor<fp16, [1, 16, 1024]> input_151_cast_fp16 = mul(x = layers_18_post_attention_layernorm_weight_to_fp16, y = hidden_states_299_cast_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_18_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_18_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(637647808)))];
            tensor<fp16, [1, 16, 4096]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_18_mlp_gate_proj_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_6838_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_6838_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_18_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_18_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646036480)))];
            tensor<fp16, [1, 16, 4096]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_18_mlp_up_proj_weight_to_fp16, x = input_151_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_155_cast_fp16 = mul(x = var_6838_cast_fp16, y = linear_131_cast_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_18_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_18_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654425152)))];
            tensor<fp16, [1, 16, 1024]> linear_132_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_18_mlp_down_proj_weight_to_fp16, x = input_155_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_303_cast_fp16 = add(x = hidden_states_295_cast_fp16, y = linear_132_cast_fp16)[name = string("hidden_states_303_cast_fp16")];
            tensor<int32, [4]> lkc_77_begin_0 = const()[name = string("lkc_77_begin_0"), val = tensor<int32, [4]>([0, 19456, 0, 0])];
            tensor<int32, [4]> lkc_77_end_0 = const()[name = string("lkc_77_end_0"), val = tensor<int32, [4]>([1, 20480, 1, 1024])];
            tensor<bool, [4]> lkc_77_end_mask_0 = const()[name = string("lkc_77_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_77_cast_fp16 = slice_by_index(begin = lkc_77_begin_0, end = lkc_77_end_0, end_mask = lkc_77_end_mask_0, x = key_cache)[name = string("lkc_77_cast_fp16")];
            tensor<int32, [4]> lvc_77_begin_0 = const()[name = string("lvc_77_begin_0"), val = tensor<int32, [4]>([0, 19456, 0, 0])];
            tensor<int32, [4]> lvc_77_end_0 = const()[name = string("lvc_77_end_0"), val = tensor<int32, [4]>([1, 20480, 1, 1024])];
            tensor<bool, [4]> lvc_77_end_mask_0 = const()[name = string("lvc_77_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_77_cast_fp16 = slice_by_index(begin = lvc_77_begin_0, end = lvc_77_end_0, end_mask = lvc_77_end_mask_0, x = value_cache)[name = string("lvc_77_cast_fp16")];
            fp16 var_6869_promoted_to_fp16 = const()[name = string("op_6869_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6875_cast_fp16 = pow(x = hidden_states_303_cast_fp16, y = var_6869_promoted_to_fp16)[name = string("op_6875_cast_fp16")];
            tensor<int32, [1]> variance_77_axes_0 = const()[name = string("variance_77_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_77_keep_dims_0 = const()[name = string("variance_77_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_77_cast_fp16 = reduce_mean(axes = variance_77_axes_0, keep_dims = variance_77_keep_dims_0, x = var_6875_cast_fp16)[name = string("variance_77_cast_fp16")];
            fp16 var_6878_to_fp16 = const()[name = string("op_6878_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_6879_cast_fp16 = add(x = variance_77_cast_fp16, y = var_6878_to_fp16)[name = string("op_6879_cast_fp16")];
            fp32 var_6880_epsilon_0 = const()[name = string("op_6880_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6880_cast_fp16 = rsqrt(epsilon = var_6880_epsilon_0, x = var_6879_cast_fp16)[name = string("op_6880_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_307_cast_fp16 = mul(x = hidden_states_303_cast_fp16, y = var_6880_cast_fp16)[name = string("hidden_states_307_cast_fp16")];
            tensor<fp16, [1024]> layers_19_input_layernorm_weight_to_fp16 = const()[name = string("layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662813824)))];
            tensor<fp16, [1, 16, 1024]> hidden_77_cast_fp16 = mul(x = layers_19_input_layernorm_weight_to_fp16, y = hidden_states_307_cast_fp16)[name = string("hidden_77_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662815936)))];
            tensor<fp16, [1, 16, 1024]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_q_proj_weight_to_fp16, x = hidden_77_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_6904 = const()[name = string("op_6904"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_115_cast_fp16 = reshape(shape = var_6904, x = linear_133_cast_fp16)[name = string("q_115_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664913152)))];
            tensor<fp16, [1, 16, 1024]> linear_134_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_k_proj_weight_to_fp16, x = hidden_77_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_6911 = const()[name = string("op_6911"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_77_cast_fp16 = reshape(shape = var_6911, x = linear_134_cast_fp16)[name = string("k_77_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667010368)))];
            tensor<fp16, [1, 16, 1024]> linear_135_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_v_proj_weight_to_fp16, x = hidden_77_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_6918 = const()[name = string("op_6918"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_39_cast_fp16 = reshape(shape = var_6918, x = linear_135_cast_fp16)[name = string("v_39_cast_fp16")];
            tensor<int32, [4]> q_117_perm_0 = const()[name = string("q_117_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_79_perm_0 = const()[name = string("k_79_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_19_perm_0 = const()[name = string("transpose_19_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_117_cast_fp16 = transpose(perm = q_117_perm_0, x = q_115_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 64]> var_6931_cast_fp16 = mul(x = q_117_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6931_cast_fp16")];
            tensor<int32, [4]> x1_77_begin_0 = const()[name = string("x1_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_77_end_0 = const()[name = string("x1_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_77_end_mask_0 = const()[name = string("x1_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_77_cast_fp16 = slice_by_index(begin = x1_77_begin_0, end = x1_77_end_0, end_mask = x1_77_end_mask_0, x = q_117_cast_fp16)[name = string("x1_77_cast_fp16")];
            tensor<int32, [4]> x2_77_begin_0 = const()[name = string("x2_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_77_end_0 = const()[name = string("x2_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_77_end_mask_0 = const()[name = string("x2_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_77_cast_fp16 = slice_by_index(begin = x2_77_begin_0, end = x2_77_end_0, end_mask = x2_77_end_mask_0, x = q_117_cast_fp16)[name = string("x2_77_cast_fp16")];
            fp16 const_235_promoted_to_fp16 = const()[name = string("const_235_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6952_cast_fp16 = mul(x = x2_77_cast_fp16, y = const_235_promoted_to_fp16)[name = string("op_6952_cast_fp16")];
            int32 var_6954 = const()[name = string("op_6954"), val = int32(-1)];
            bool var_6955_interleave_0 = const()[name = string("op_6955_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6955_cast_fp16 = concat(axis = var_6954, interleave = var_6955_interleave_0, values = (var_6952_cast_fp16, x1_77_cast_fp16))[name = string("op_6955_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6958_cast_fp16 = mul(x = var_6955_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_119_cast_fp16 = add(x = var_6931_cast_fp16, y = var_6958_cast_fp16)[name = string("q_119_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_79_cast_fp16 = transpose(perm = k_79_perm_0, x = k_77_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 16, 16, 64]> var_6963_cast_fp16 = mul(x = k_79_cast_fp16, y = var_1078_cast_fp16)[name = string("op_6963_cast_fp16")];
            tensor<int32, [4]> x1_79_begin_0 = const()[name = string("x1_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_79_end_0 = const()[name = string("x1_79_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_79_end_mask_0 = const()[name = string("x1_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_79_cast_fp16 = slice_by_index(begin = x1_79_begin_0, end = x1_79_end_0, end_mask = x1_79_end_mask_0, x = k_79_cast_fp16)[name = string("x1_79_cast_fp16")];
            tensor<int32, [4]> x2_79_begin_0 = const()[name = string("x2_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_79_end_0 = const()[name = string("x2_79_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_79_end_mask_0 = const()[name = string("x2_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_79_cast_fp16 = slice_by_index(begin = x2_79_begin_0, end = x2_79_end_0, end_mask = x2_79_end_mask_0, x = k_79_cast_fp16)[name = string("x2_79_cast_fp16")];
            fp16 const_238_promoted_to_fp16 = const()[name = string("const_238_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_6984_cast_fp16 = mul(x = x2_79_cast_fp16, y = const_238_promoted_to_fp16)[name = string("op_6984_cast_fp16")];
            int32 var_6986 = const()[name = string("op_6986"), val = int32(-1)];
            bool var_6987_interleave_0 = const()[name = string("op_6987_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_6987_cast_fp16 = concat(axis = var_6986, interleave = var_6987_interleave_0, values = (var_6984_cast_fp16, x1_79_cast_fp16))[name = string("op_6987_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6990_cast_fp16 = mul(x = var_6987_cast_fp16, y = var_1105_cast_fp16)[name = string("op_6990_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_77_cast_fp16 = add(x = var_6963_cast_fp16, y = var_6990_cast_fp16)[name = string("x_77_cast_fp16")];
            tensor<int32, [4]> var_6997 = const()[name = string("op_6997"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7009 = const()[name = string("op_7009"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_6998_cast_fp16 = transpose(perm = var_6997, x = x_77_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 1024, 1, 16]> var_7010_cast_fp16 = reshape(shape = var_7009, x = var_6998_cast_fp16)[name = string("op_7010_cast_fp16")];
            tensor<int32, [1]> var_7012_axes_0 = const()[name = string("op_7012_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7012_cast_fp16 = squeeze(axes = var_7012_axes_0, x = var_7010_cast_fp16)[name = string("op_7012_cast_fp16")];
            bool var_7013_transpose_x_0 = const()[name = string("op_7013_transpose_x_0"), val = bool(false)];
            bool var_7013_transpose_y_0 = const()[name = string("op_7013_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7013_cast_fp16 = matmul(transpose_x = var_7013_transpose_x_0, transpose_y = var_7013_transpose_y_0, x = var_7012_cast_fp16, y = block_cache_map)[name = string("op_7013_cast_fp16")];
            tensor<int32, [1]> nk_full_39_axes_0 = const()[name = string("nk_full_39_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_39_cast_fp16 = expand_dims(axes = nk_full_39_axes_0, x = var_7013_cast_fp16)[name = string("nk_full_39_cast_fp16")];
            tensor<int32, [4]> var_7032 = const()[name = string("op_7032"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_19_cast_fp16 = transpose(perm = transpose_19_perm_0, x = v_39_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 1024, 1, 16]> var_7033_cast_fp16 = reshape(shape = var_7032, x = transpose_19_cast_fp16)[name = string("op_7033_cast_fp16")];
            tensor<int32, [1]> var_7035_axes_0 = const()[name = string("op_7035_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7035_cast_fp16 = squeeze(axes = var_7035_axes_0, x = var_7033_cast_fp16)[name = string("op_7035_cast_fp16")];
            bool var_7036_transpose_x_0 = const()[name = string("op_7036_transpose_x_0"), val = bool(false)];
            bool var_7036_transpose_y_0 = const()[name = string("op_7036_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7036_cast_fp16 = matmul(transpose_x = var_7036_transpose_x_0, transpose_y = var_7036_transpose_y_0, x = var_7035_cast_fp16, y = block_cache_map)[name = string("op_7036_cast_fp16")];
            tensor<int32, [1]> nv_full_39_axes_0 = const()[name = string("nv_full_39_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_39_cast_fp16 = expand_dims(axes = nv_full_39_axes_0, x = var_7036_cast_fp16)[name = string("nv_full_39_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7042_cast_fp16 = mul(x = lkc_77_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7042_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7043_cast_fp16 = mul(x = nk_full_39_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7043_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_79_cast_fp16 = add(x = var_7042_cast_fp16, y = var_7043_cast_fp16)[name = string("lkc_79_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7049_cast_fp16 = mul(x = lvc_77_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7049_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7050_cast_fp16 = mul(x = nv_full_39_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7050_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_79_cast_fp16 = add(x = var_7049_cast_fp16, y = var_7050_cast_fp16)[name = string("lvc_79_cast_fp16")];
            tensor<int32, [1]> var_7054_axes_0 = const()[name = string("op_7054_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7054_cast_fp16 = squeeze(axes = var_7054_axes_0, x = lkc_79_cast_fp16)[name = string("op_7054_cast_fp16")];
            tensor<int32, [4]> var_7058 = const()[name = string("op_7058"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_39_cast_fp16 = reshape(shape = var_7058, x = var_7054_cast_fp16)[name = string("kc_39_cast_fp16")];
            tensor<int32, [1]> var_7061_axes_0 = const()[name = string("op_7061_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7061_cast_fp16 = squeeze(axes = var_7061_axes_0, x = lvc_79_cast_fp16)[name = string("op_7061_cast_fp16")];
            tensor<int32, [4]> var_7065 = const()[name = string("op_7065"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_39_cast_fp16 = reshape(shape = var_7065, x = var_7061_cast_fp16)[name = string("vc_39_cast_fp16")];
            bool var_7067_transpose_x_0 = const()[name = string("op_7067_transpose_x_0"), val = bool(false)];
            bool var_7067_transpose_y_0 = const()[name = string("op_7067_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_7067_cast_fp16 = matmul(transpose_x = var_7067_transpose_x_0, transpose_y = var_7067_transpose_y_0, x = q_119_cast_fp16, y = kc_39_cast_fp16)[name = string("op_7067_cast_fp16")];
            fp16 _inversed_scores_115_y_0_to_fp16 = const()[name = string("_inversed_scores_115_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_115_cast_fp16 = mul(x = var_7067_cast_fp16, y = _inversed_scores_115_y_0_to_fp16)[name = string("_inversed_scores_115_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_117_cast_fp16 = add(x = _inversed_scores_115_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_117_cast_fp16")];
            int32 var_7096 = const()[name = string("op_7096"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_7098_cast_fp16 = softmax(axis = var_7096, x = scores_117_cast_fp16)[name = string("op_7098_cast_fp16")];
            bool attn_out_77_transpose_x_1 = const()[name = string("attn_out_77_transpose_x_1"), val = bool(false)];
            bool attn_out_77_transpose_y_1 = const()[name = string("attn_out_77_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_77_cast_fp16 = matmul(transpose_x = attn_out_77_transpose_x_1, transpose_y = attn_out_77_transpose_y_1, x = var_7098_cast_fp16, y = vc_39_cast_fp16)[name = string("attn_out_77_cast_fp16")];
            tensor<int32, [4]> var_7110_perm_0 = const()[name = string("op_7110_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7112 = const()[name = string("op_7112"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_7110_cast_fp16 = transpose(perm = var_7110_perm_0, x = attn_out_77_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 1024]> input_157_cast_fp16 = reshape(shape = var_7112, x = var_7110_cast_fp16)[name = string("input_157_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_19_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_19_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(669107584)))];
            tensor<fp16, [1, 16, 1024]> linear_136_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_self_attn_o_proj_weight_to_fp16, x = input_157_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_311_cast_fp16 = add(x = hidden_states_303_cast_fp16, y = linear_136_cast_fp16)[name = string("hidden_states_311_cast_fp16")];
            fp16 var_7122_promoted_to_fp16 = const()[name = string("op_7122_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7128_cast_fp16 = pow(x = hidden_states_311_cast_fp16, y = var_7122_promoted_to_fp16)[name = string("op_7128_cast_fp16")];
            tensor<int32, [1]> variance_79_axes_0 = const()[name = string("variance_79_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_79_keep_dims_0 = const()[name = string("variance_79_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_79_cast_fp16 = reduce_mean(axes = variance_79_axes_0, keep_dims = variance_79_keep_dims_0, x = var_7128_cast_fp16)[name = string("variance_79_cast_fp16")];
            fp16 var_7131_to_fp16 = const()[name = string("op_7131_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7132_cast_fp16 = add(x = variance_79_cast_fp16, y = var_7131_to_fp16)[name = string("op_7132_cast_fp16")];
            fp32 var_7133_epsilon_0 = const()[name = string("op_7133_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7133_cast_fp16 = rsqrt(epsilon = var_7133_epsilon_0, x = var_7132_cast_fp16)[name = string("op_7133_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_315_cast_fp16 = mul(x = hidden_states_311_cast_fp16, y = var_7133_cast_fp16)[name = string("hidden_states_315_cast_fp16")];
            tensor<fp16, [1024]> layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671204800)))];
            tensor<fp16, [1, 16, 1024]> input_159_cast_fp16 = mul(x = layers_19_post_attention_layernorm_weight_to_fp16, y = hidden_states_315_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_19_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_19_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(671206912)))];
            tensor<fp16, [1, 16, 4096]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_19_mlp_gate_proj_weight_to_fp16, x = input_159_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_7146_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7146_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_19_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_19_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(679595584)))];
            tensor<fp16, [1, 16, 4096]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_19_mlp_up_proj_weight_to_fp16, x = input_159_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_163_cast_fp16 = mul(x = var_7146_cast_fp16, y = linear_138_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_19_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_19_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(687984256)))];
            tensor<fp16, [1, 16, 1024]> linear_139_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_19_mlp_down_proj_weight_to_fp16, x = input_163_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_319_cast_fp16 = add(x = hidden_states_311_cast_fp16, y = linear_139_cast_fp16)[name = string("hidden_states_319_cast_fp16")];
            tensor<int32, [4]> lkc_81_begin_0 = const()[name = string("lkc_81_begin_0"), val = tensor<int32, [4]>([0, 20480, 0, 0])];
            tensor<int32, [4]> lkc_81_end_0 = const()[name = string("lkc_81_end_0"), val = tensor<int32, [4]>([1, 21504, 1, 1024])];
            tensor<bool, [4]> lkc_81_end_mask_0 = const()[name = string("lkc_81_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_81_cast_fp16 = slice_by_index(begin = lkc_81_begin_0, end = lkc_81_end_0, end_mask = lkc_81_end_mask_0, x = key_cache)[name = string("lkc_81_cast_fp16")];
            tensor<int32, [4]> lvc_81_begin_0 = const()[name = string("lvc_81_begin_0"), val = tensor<int32, [4]>([0, 20480, 0, 0])];
            tensor<int32, [4]> lvc_81_end_0 = const()[name = string("lvc_81_end_0"), val = tensor<int32, [4]>([1, 21504, 1, 1024])];
            tensor<bool, [4]> lvc_81_end_mask_0 = const()[name = string("lvc_81_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_81_cast_fp16 = slice_by_index(begin = lvc_81_begin_0, end = lvc_81_end_0, end_mask = lvc_81_end_mask_0, x = value_cache)[name = string("lvc_81_cast_fp16")];
            fp16 var_7177_promoted_to_fp16 = const()[name = string("op_7177_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7183_cast_fp16 = pow(x = hidden_states_319_cast_fp16, y = var_7177_promoted_to_fp16)[name = string("op_7183_cast_fp16")];
            tensor<int32, [1]> variance_81_axes_0 = const()[name = string("variance_81_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_81_keep_dims_0 = const()[name = string("variance_81_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_81_cast_fp16 = reduce_mean(axes = variance_81_axes_0, keep_dims = variance_81_keep_dims_0, x = var_7183_cast_fp16)[name = string("variance_81_cast_fp16")];
            fp16 var_7186_to_fp16 = const()[name = string("op_7186_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7187_cast_fp16 = add(x = variance_81_cast_fp16, y = var_7186_to_fp16)[name = string("op_7187_cast_fp16")];
            fp32 var_7188_epsilon_0 = const()[name = string("op_7188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7188_cast_fp16 = rsqrt(epsilon = var_7188_epsilon_0, x = var_7187_cast_fp16)[name = string("op_7188_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_323_cast_fp16 = mul(x = hidden_states_319_cast_fp16, y = var_7188_cast_fp16)[name = string("hidden_states_323_cast_fp16")];
            tensor<fp16, [1024]> layers_20_input_layernorm_weight_to_fp16 = const()[name = string("layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696372928)))];
            tensor<fp16, [1, 16, 1024]> hidden_81_cast_fp16 = mul(x = layers_20_input_layernorm_weight_to_fp16, y = hidden_states_323_cast_fp16)[name = string("hidden_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(696375040)))];
            tensor<fp16, [1, 16, 1024]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_q_proj_weight_to_fp16, x = hidden_81_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_7212 = const()[name = string("op_7212"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_121_cast_fp16 = reshape(shape = var_7212, x = linear_140_cast_fp16)[name = string("q_121_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698472256)))];
            tensor<fp16, [1, 16, 1024]> linear_141_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_k_proj_weight_to_fp16, x = hidden_81_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_7219 = const()[name = string("op_7219"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_81_cast_fp16 = reshape(shape = var_7219, x = linear_141_cast_fp16)[name = string("k_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(700569472)))];
            tensor<fp16, [1, 16, 1024]> linear_142_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_v_proj_weight_to_fp16, x = hidden_81_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_7226 = const()[name = string("op_7226"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_41_cast_fp16 = reshape(shape = var_7226, x = linear_142_cast_fp16)[name = string("v_41_cast_fp16")];
            tensor<int32, [4]> q_123_perm_0 = const()[name = string("q_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_83_perm_0 = const()[name = string("k_83_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_20_perm_0 = const()[name = string("transpose_20_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_123_cast_fp16 = transpose(perm = q_123_perm_0, x = q_121_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 64]> var_7239_cast_fp16 = mul(x = q_123_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7239_cast_fp16")];
            tensor<int32, [4]> x1_81_begin_0 = const()[name = string("x1_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_81_end_0 = const()[name = string("x1_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_81_end_mask_0 = const()[name = string("x1_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_81_cast_fp16 = slice_by_index(begin = x1_81_begin_0, end = x1_81_end_0, end_mask = x1_81_end_mask_0, x = q_123_cast_fp16)[name = string("x1_81_cast_fp16")];
            tensor<int32, [4]> x2_81_begin_0 = const()[name = string("x2_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_81_end_0 = const()[name = string("x2_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_81_end_mask_0 = const()[name = string("x2_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_81_cast_fp16 = slice_by_index(begin = x2_81_begin_0, end = x2_81_end_0, end_mask = x2_81_end_mask_0, x = q_123_cast_fp16)[name = string("x2_81_cast_fp16")];
            fp16 const_247_promoted_to_fp16 = const()[name = string("const_247_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7260_cast_fp16 = mul(x = x2_81_cast_fp16, y = const_247_promoted_to_fp16)[name = string("op_7260_cast_fp16")];
            int32 var_7262 = const()[name = string("op_7262"), val = int32(-1)];
            bool var_7263_interleave_0 = const()[name = string("op_7263_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7263_cast_fp16 = concat(axis = var_7262, interleave = var_7263_interleave_0, values = (var_7260_cast_fp16, x1_81_cast_fp16))[name = string("op_7263_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7266_cast_fp16 = mul(x = var_7263_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7266_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_125_cast_fp16 = add(x = var_7239_cast_fp16, y = var_7266_cast_fp16)[name = string("q_125_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_83_cast_fp16 = transpose(perm = k_83_perm_0, x = k_81_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 16, 16, 64]> var_7271_cast_fp16 = mul(x = k_83_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7271_cast_fp16")];
            tensor<int32, [4]> x1_83_begin_0 = const()[name = string("x1_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_83_end_0 = const()[name = string("x1_83_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_83_end_mask_0 = const()[name = string("x1_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_83_cast_fp16 = slice_by_index(begin = x1_83_begin_0, end = x1_83_end_0, end_mask = x1_83_end_mask_0, x = k_83_cast_fp16)[name = string("x1_83_cast_fp16")];
            tensor<int32, [4]> x2_83_begin_0 = const()[name = string("x2_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_83_end_0 = const()[name = string("x2_83_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_83_end_mask_0 = const()[name = string("x2_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_83_cast_fp16 = slice_by_index(begin = x2_83_begin_0, end = x2_83_end_0, end_mask = x2_83_end_mask_0, x = k_83_cast_fp16)[name = string("x2_83_cast_fp16")];
            fp16 const_250_promoted_to_fp16 = const()[name = string("const_250_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7292_cast_fp16 = mul(x = x2_83_cast_fp16, y = const_250_promoted_to_fp16)[name = string("op_7292_cast_fp16")];
            int32 var_7294 = const()[name = string("op_7294"), val = int32(-1)];
            bool var_7295_interleave_0 = const()[name = string("op_7295_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7295_cast_fp16 = concat(axis = var_7294, interleave = var_7295_interleave_0, values = (var_7292_cast_fp16, x1_83_cast_fp16))[name = string("op_7295_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7298_cast_fp16 = mul(x = var_7295_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7298_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_81_cast_fp16 = add(x = var_7271_cast_fp16, y = var_7298_cast_fp16)[name = string("x_81_cast_fp16")];
            tensor<int32, [4]> var_7305 = const()[name = string("op_7305"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7317 = const()[name = string("op_7317"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_7306_cast_fp16 = transpose(perm = var_7305, x = x_81_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 1024, 1, 16]> var_7318_cast_fp16 = reshape(shape = var_7317, x = var_7306_cast_fp16)[name = string("op_7318_cast_fp16")];
            tensor<int32, [1]> var_7320_axes_0 = const()[name = string("op_7320_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7320_cast_fp16 = squeeze(axes = var_7320_axes_0, x = var_7318_cast_fp16)[name = string("op_7320_cast_fp16")];
            bool var_7321_transpose_x_0 = const()[name = string("op_7321_transpose_x_0"), val = bool(false)];
            bool var_7321_transpose_y_0 = const()[name = string("op_7321_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7321_cast_fp16 = matmul(transpose_x = var_7321_transpose_x_0, transpose_y = var_7321_transpose_y_0, x = var_7320_cast_fp16, y = block_cache_map)[name = string("op_7321_cast_fp16")];
            tensor<int32, [1]> nk_full_41_axes_0 = const()[name = string("nk_full_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_41_cast_fp16 = expand_dims(axes = nk_full_41_axes_0, x = var_7321_cast_fp16)[name = string("nk_full_41_cast_fp16")];
            tensor<int32, [4]> var_7340 = const()[name = string("op_7340"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_20_cast_fp16 = transpose(perm = transpose_20_perm_0, x = v_41_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 1024, 1, 16]> var_7341_cast_fp16 = reshape(shape = var_7340, x = transpose_20_cast_fp16)[name = string("op_7341_cast_fp16")];
            tensor<int32, [1]> var_7343_axes_0 = const()[name = string("op_7343_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7343_cast_fp16 = squeeze(axes = var_7343_axes_0, x = var_7341_cast_fp16)[name = string("op_7343_cast_fp16")];
            bool var_7344_transpose_x_0 = const()[name = string("op_7344_transpose_x_0"), val = bool(false)];
            bool var_7344_transpose_y_0 = const()[name = string("op_7344_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7344_cast_fp16 = matmul(transpose_x = var_7344_transpose_x_0, transpose_y = var_7344_transpose_y_0, x = var_7343_cast_fp16, y = block_cache_map)[name = string("op_7344_cast_fp16")];
            tensor<int32, [1]> nv_full_41_axes_0 = const()[name = string("nv_full_41_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_41_cast_fp16 = expand_dims(axes = nv_full_41_axes_0, x = var_7344_cast_fp16)[name = string("nv_full_41_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7350_cast_fp16 = mul(x = lkc_81_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7350_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7351_cast_fp16 = mul(x = nk_full_41_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7351_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_83_cast_fp16 = add(x = var_7350_cast_fp16, y = var_7351_cast_fp16)[name = string("lkc_83_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7357_cast_fp16 = mul(x = lvc_81_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7357_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7358_cast_fp16 = mul(x = nv_full_41_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7358_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_83_cast_fp16 = add(x = var_7357_cast_fp16, y = var_7358_cast_fp16)[name = string("lvc_83_cast_fp16")];
            tensor<int32, [1]> var_7362_axes_0 = const()[name = string("op_7362_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7362_cast_fp16 = squeeze(axes = var_7362_axes_0, x = lkc_83_cast_fp16)[name = string("op_7362_cast_fp16")];
            tensor<int32, [4]> var_7366 = const()[name = string("op_7366"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_41_cast_fp16 = reshape(shape = var_7366, x = var_7362_cast_fp16)[name = string("kc_41_cast_fp16")];
            tensor<int32, [1]> var_7369_axes_0 = const()[name = string("op_7369_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7369_cast_fp16 = squeeze(axes = var_7369_axes_0, x = lvc_83_cast_fp16)[name = string("op_7369_cast_fp16")];
            tensor<int32, [4]> var_7373 = const()[name = string("op_7373"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_41_cast_fp16 = reshape(shape = var_7373, x = var_7369_cast_fp16)[name = string("vc_41_cast_fp16")];
            bool var_7375_transpose_x_0 = const()[name = string("op_7375_transpose_x_0"), val = bool(false)];
            bool var_7375_transpose_y_0 = const()[name = string("op_7375_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_7375_cast_fp16 = matmul(transpose_x = var_7375_transpose_x_0, transpose_y = var_7375_transpose_y_0, x = q_125_cast_fp16, y = kc_41_cast_fp16)[name = string("op_7375_cast_fp16")];
            fp16 _inversed_scores_121_y_0_to_fp16 = const()[name = string("_inversed_scores_121_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_121_cast_fp16 = mul(x = var_7375_cast_fp16, y = _inversed_scores_121_y_0_to_fp16)[name = string("_inversed_scores_121_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_123_cast_fp16 = add(x = _inversed_scores_121_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_123_cast_fp16")];
            int32 var_7404 = const()[name = string("op_7404"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_7406_cast_fp16 = softmax(axis = var_7404, x = scores_123_cast_fp16)[name = string("op_7406_cast_fp16")];
            bool attn_out_81_transpose_x_1 = const()[name = string("attn_out_81_transpose_x_1"), val = bool(false)];
            bool attn_out_81_transpose_y_1 = const()[name = string("attn_out_81_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_81_cast_fp16 = matmul(transpose_x = attn_out_81_transpose_x_1, transpose_y = attn_out_81_transpose_y_1, x = var_7406_cast_fp16, y = vc_41_cast_fp16)[name = string("attn_out_81_cast_fp16")];
            tensor<int32, [4]> var_7418_perm_0 = const()[name = string("op_7418_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7420 = const()[name = string("op_7420"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_7418_cast_fp16 = transpose(perm = var_7418_perm_0, x = attn_out_81_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 1024]> input_165_cast_fp16 = reshape(shape = var_7420, x = var_7418_cast_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_20_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_20_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702666688)))];
            tensor<fp16, [1, 16, 1024]> linear_143_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_self_attn_o_proj_weight_to_fp16, x = input_165_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_327_cast_fp16 = add(x = hidden_states_319_cast_fp16, y = linear_143_cast_fp16)[name = string("hidden_states_327_cast_fp16")];
            fp16 var_7430_promoted_to_fp16 = const()[name = string("op_7430_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7436_cast_fp16 = pow(x = hidden_states_327_cast_fp16, y = var_7430_promoted_to_fp16)[name = string("op_7436_cast_fp16")];
            tensor<int32, [1]> variance_83_axes_0 = const()[name = string("variance_83_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_83_keep_dims_0 = const()[name = string("variance_83_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_83_cast_fp16 = reduce_mean(axes = variance_83_axes_0, keep_dims = variance_83_keep_dims_0, x = var_7436_cast_fp16)[name = string("variance_83_cast_fp16")];
            fp16 var_7439_to_fp16 = const()[name = string("op_7439_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7440_cast_fp16 = add(x = variance_83_cast_fp16, y = var_7439_to_fp16)[name = string("op_7440_cast_fp16")];
            fp32 var_7441_epsilon_0 = const()[name = string("op_7441_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7441_cast_fp16 = rsqrt(epsilon = var_7441_epsilon_0, x = var_7440_cast_fp16)[name = string("op_7441_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_331_cast_fp16 = mul(x = hidden_states_327_cast_fp16, y = var_7441_cast_fp16)[name = string("hidden_states_331_cast_fp16")];
            tensor<fp16, [1024]> layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704763904)))];
            tensor<fp16, [1, 16, 1024]> input_167_cast_fp16 = mul(x = layers_20_post_attention_layernorm_weight_to_fp16, y = hidden_states_331_cast_fp16)[name = string("input_167_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_20_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_20_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704766016)))];
            tensor<fp16, [1, 16, 4096]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_20_mlp_gate_proj_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_7454_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_7454_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_20_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_20_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(713154688)))];
            tensor<fp16, [1, 16, 4096]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_20_mlp_up_proj_weight_to_fp16, x = input_167_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_171_cast_fp16 = mul(x = var_7454_cast_fp16, y = linear_145_cast_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_20_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_20_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(721543360)))];
            tensor<fp16, [1, 16, 1024]> linear_146_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_20_mlp_down_proj_weight_to_fp16, x = input_171_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_335_cast_fp16 = add(x = hidden_states_327_cast_fp16, y = linear_146_cast_fp16)[name = string("hidden_states_335_cast_fp16")];
            tensor<int32, [4]> lkc_85_begin_0 = const()[name = string("lkc_85_begin_0"), val = tensor<int32, [4]>([0, 21504, 0, 0])];
            tensor<int32, [4]> lkc_85_end_0 = const()[name = string("lkc_85_end_0"), val = tensor<int32, [4]>([1, 22528, 1, 1024])];
            tensor<bool, [4]> lkc_85_end_mask_0 = const()[name = string("lkc_85_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_85_cast_fp16 = slice_by_index(begin = lkc_85_begin_0, end = lkc_85_end_0, end_mask = lkc_85_end_mask_0, x = key_cache)[name = string("lkc_85_cast_fp16")];
            tensor<int32, [4]> lvc_85_begin_0 = const()[name = string("lvc_85_begin_0"), val = tensor<int32, [4]>([0, 21504, 0, 0])];
            tensor<int32, [4]> lvc_85_end_0 = const()[name = string("lvc_85_end_0"), val = tensor<int32, [4]>([1, 22528, 1, 1024])];
            tensor<bool, [4]> lvc_85_end_mask_0 = const()[name = string("lvc_85_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_85_cast_fp16 = slice_by_index(begin = lvc_85_begin_0, end = lvc_85_end_0, end_mask = lvc_85_end_mask_0, x = value_cache)[name = string("lvc_85_cast_fp16")];
            fp16 var_7485_promoted_to_fp16 = const()[name = string("op_7485_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7491_cast_fp16 = pow(x = hidden_states_335_cast_fp16, y = var_7485_promoted_to_fp16)[name = string("op_7491_cast_fp16")];
            tensor<int32, [1]> variance_85_axes_0 = const()[name = string("variance_85_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_85_keep_dims_0 = const()[name = string("variance_85_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_85_cast_fp16 = reduce_mean(axes = variance_85_axes_0, keep_dims = variance_85_keep_dims_0, x = var_7491_cast_fp16)[name = string("variance_85_cast_fp16")];
            fp16 var_7494_to_fp16 = const()[name = string("op_7494_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7495_cast_fp16 = add(x = variance_85_cast_fp16, y = var_7494_to_fp16)[name = string("op_7495_cast_fp16")];
            fp32 var_7496_epsilon_0 = const()[name = string("op_7496_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7496_cast_fp16 = rsqrt(epsilon = var_7496_epsilon_0, x = var_7495_cast_fp16)[name = string("op_7496_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_339_cast_fp16 = mul(x = hidden_states_335_cast_fp16, y = var_7496_cast_fp16)[name = string("hidden_states_339_cast_fp16")];
            tensor<fp16, [1024]> layers_21_input_layernorm_weight_to_fp16 = const()[name = string("layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729932032)))];
            tensor<fp16, [1, 16, 1024]> hidden_85_cast_fp16 = mul(x = layers_21_input_layernorm_weight_to_fp16, y = hidden_states_339_cast_fp16)[name = string("hidden_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(729934144)))];
            tensor<fp16, [1, 16, 1024]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_q_proj_weight_to_fp16, x = hidden_85_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7520 = const()[name = string("op_7520"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_127_cast_fp16 = reshape(shape = var_7520, x = linear_147_cast_fp16)[name = string("q_127_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(732031360)))];
            tensor<fp16, [1, 16, 1024]> linear_148_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_k_proj_weight_to_fp16, x = hidden_85_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_7527 = const()[name = string("op_7527"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_85_cast_fp16 = reshape(shape = var_7527, x = linear_148_cast_fp16)[name = string("k_85_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(734128576)))];
            tensor<fp16, [1, 16, 1024]> linear_149_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_v_proj_weight_to_fp16, x = hidden_85_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_7534 = const()[name = string("op_7534"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_43_cast_fp16 = reshape(shape = var_7534, x = linear_149_cast_fp16)[name = string("v_43_cast_fp16")];
            tensor<int32, [4]> q_129_perm_0 = const()[name = string("q_129_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_87_perm_0 = const()[name = string("k_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_21_perm_0 = const()[name = string("transpose_21_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_129_cast_fp16 = transpose(perm = q_129_perm_0, x = q_127_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 64]> var_7547_cast_fp16 = mul(x = q_129_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7547_cast_fp16")];
            tensor<int32, [4]> x1_85_begin_0 = const()[name = string("x1_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_85_end_0 = const()[name = string("x1_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_85_end_mask_0 = const()[name = string("x1_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_85_cast_fp16 = slice_by_index(begin = x1_85_begin_0, end = x1_85_end_0, end_mask = x1_85_end_mask_0, x = q_129_cast_fp16)[name = string("x1_85_cast_fp16")];
            tensor<int32, [4]> x2_85_begin_0 = const()[name = string("x2_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_85_end_0 = const()[name = string("x2_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_85_end_mask_0 = const()[name = string("x2_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_85_cast_fp16 = slice_by_index(begin = x2_85_begin_0, end = x2_85_end_0, end_mask = x2_85_end_mask_0, x = q_129_cast_fp16)[name = string("x2_85_cast_fp16")];
            fp16 const_259_promoted_to_fp16 = const()[name = string("const_259_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7568_cast_fp16 = mul(x = x2_85_cast_fp16, y = const_259_promoted_to_fp16)[name = string("op_7568_cast_fp16")];
            int32 var_7570 = const()[name = string("op_7570"), val = int32(-1)];
            bool var_7571_interleave_0 = const()[name = string("op_7571_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7571_cast_fp16 = concat(axis = var_7570, interleave = var_7571_interleave_0, values = (var_7568_cast_fp16, x1_85_cast_fp16))[name = string("op_7571_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7574_cast_fp16 = mul(x = var_7571_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7574_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_131_cast_fp16 = add(x = var_7547_cast_fp16, y = var_7574_cast_fp16)[name = string("q_131_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_87_cast_fp16 = transpose(perm = k_87_perm_0, x = k_85_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 16, 16, 64]> var_7579_cast_fp16 = mul(x = k_87_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7579_cast_fp16")];
            tensor<int32, [4]> x1_87_begin_0 = const()[name = string("x1_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_87_end_0 = const()[name = string("x1_87_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_87_end_mask_0 = const()[name = string("x1_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_87_cast_fp16 = slice_by_index(begin = x1_87_begin_0, end = x1_87_end_0, end_mask = x1_87_end_mask_0, x = k_87_cast_fp16)[name = string("x1_87_cast_fp16")];
            tensor<int32, [4]> x2_87_begin_0 = const()[name = string("x2_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_87_end_0 = const()[name = string("x2_87_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_87_end_mask_0 = const()[name = string("x2_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_87_cast_fp16 = slice_by_index(begin = x2_87_begin_0, end = x2_87_end_0, end_mask = x2_87_end_mask_0, x = k_87_cast_fp16)[name = string("x2_87_cast_fp16")];
            fp16 const_262_promoted_to_fp16 = const()[name = string("const_262_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7600_cast_fp16 = mul(x = x2_87_cast_fp16, y = const_262_promoted_to_fp16)[name = string("op_7600_cast_fp16")];
            int32 var_7602 = const()[name = string("op_7602"), val = int32(-1)];
            bool var_7603_interleave_0 = const()[name = string("op_7603_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7603_cast_fp16 = concat(axis = var_7602, interleave = var_7603_interleave_0, values = (var_7600_cast_fp16, x1_87_cast_fp16))[name = string("op_7603_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7606_cast_fp16 = mul(x = var_7603_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7606_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_85_cast_fp16 = add(x = var_7579_cast_fp16, y = var_7606_cast_fp16)[name = string("x_85_cast_fp16")];
            tensor<int32, [4]> var_7613 = const()[name = string("op_7613"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7625 = const()[name = string("op_7625"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_7614_cast_fp16 = transpose(perm = var_7613, x = x_85_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 1024, 1, 16]> var_7626_cast_fp16 = reshape(shape = var_7625, x = var_7614_cast_fp16)[name = string("op_7626_cast_fp16")];
            tensor<int32, [1]> var_7628_axes_0 = const()[name = string("op_7628_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7628_cast_fp16 = squeeze(axes = var_7628_axes_0, x = var_7626_cast_fp16)[name = string("op_7628_cast_fp16")];
            bool var_7629_transpose_x_0 = const()[name = string("op_7629_transpose_x_0"), val = bool(false)];
            bool var_7629_transpose_y_0 = const()[name = string("op_7629_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7629_cast_fp16 = matmul(transpose_x = var_7629_transpose_x_0, transpose_y = var_7629_transpose_y_0, x = var_7628_cast_fp16, y = block_cache_map)[name = string("op_7629_cast_fp16")];
            tensor<int32, [1]> nk_full_43_axes_0 = const()[name = string("nk_full_43_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_43_cast_fp16 = expand_dims(axes = nk_full_43_axes_0, x = var_7629_cast_fp16)[name = string("nk_full_43_cast_fp16")];
            tensor<int32, [4]> var_7648 = const()[name = string("op_7648"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_21_cast_fp16 = transpose(perm = transpose_21_perm_0, x = v_43_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 1024, 1, 16]> var_7649_cast_fp16 = reshape(shape = var_7648, x = transpose_21_cast_fp16)[name = string("op_7649_cast_fp16")];
            tensor<int32, [1]> var_7651_axes_0 = const()[name = string("op_7651_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7651_cast_fp16 = squeeze(axes = var_7651_axes_0, x = var_7649_cast_fp16)[name = string("op_7651_cast_fp16")];
            bool var_7652_transpose_x_0 = const()[name = string("op_7652_transpose_x_0"), val = bool(false)];
            bool var_7652_transpose_y_0 = const()[name = string("op_7652_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7652_cast_fp16 = matmul(transpose_x = var_7652_transpose_x_0, transpose_y = var_7652_transpose_y_0, x = var_7651_cast_fp16, y = block_cache_map)[name = string("op_7652_cast_fp16")];
            tensor<int32, [1]> nv_full_43_axes_0 = const()[name = string("nv_full_43_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_43_cast_fp16 = expand_dims(axes = nv_full_43_axes_0, x = var_7652_cast_fp16)[name = string("nv_full_43_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7658_cast_fp16 = mul(x = lkc_85_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7658_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7659_cast_fp16 = mul(x = nk_full_43_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7659_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_87_cast_fp16 = add(x = var_7658_cast_fp16, y = var_7659_cast_fp16)[name = string("lkc_87_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7665_cast_fp16 = mul(x = lvc_85_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7665_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7666_cast_fp16 = mul(x = nv_full_43_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7666_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_87_cast_fp16 = add(x = var_7665_cast_fp16, y = var_7666_cast_fp16)[name = string("lvc_87_cast_fp16")];
            tensor<int32, [1]> var_7670_axes_0 = const()[name = string("op_7670_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7670_cast_fp16 = squeeze(axes = var_7670_axes_0, x = lkc_87_cast_fp16)[name = string("op_7670_cast_fp16")];
            tensor<int32, [4]> var_7674 = const()[name = string("op_7674"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_43_cast_fp16 = reshape(shape = var_7674, x = var_7670_cast_fp16)[name = string("kc_43_cast_fp16")];
            tensor<int32, [1]> var_7677_axes_0 = const()[name = string("op_7677_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7677_cast_fp16 = squeeze(axes = var_7677_axes_0, x = lvc_87_cast_fp16)[name = string("op_7677_cast_fp16")];
            tensor<int32, [4]> var_7681 = const()[name = string("op_7681"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_43_cast_fp16 = reshape(shape = var_7681, x = var_7677_cast_fp16)[name = string("vc_43_cast_fp16")];
            bool var_7683_transpose_x_0 = const()[name = string("op_7683_transpose_x_0"), val = bool(false)];
            bool var_7683_transpose_y_0 = const()[name = string("op_7683_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_7683_cast_fp16 = matmul(transpose_x = var_7683_transpose_x_0, transpose_y = var_7683_transpose_y_0, x = q_131_cast_fp16, y = kc_43_cast_fp16)[name = string("op_7683_cast_fp16")];
            fp16 _inversed_scores_127_y_0_to_fp16 = const()[name = string("_inversed_scores_127_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_127_cast_fp16 = mul(x = var_7683_cast_fp16, y = _inversed_scores_127_y_0_to_fp16)[name = string("_inversed_scores_127_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_129_cast_fp16 = add(x = _inversed_scores_127_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_129_cast_fp16")];
            int32 var_7712 = const()[name = string("op_7712"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_7714_cast_fp16 = softmax(axis = var_7712, x = scores_129_cast_fp16)[name = string("op_7714_cast_fp16")];
            bool attn_out_85_transpose_x_1 = const()[name = string("attn_out_85_transpose_x_1"), val = bool(false)];
            bool attn_out_85_transpose_y_1 = const()[name = string("attn_out_85_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_85_cast_fp16 = matmul(transpose_x = attn_out_85_transpose_x_1, transpose_y = attn_out_85_transpose_y_1, x = var_7714_cast_fp16, y = vc_43_cast_fp16)[name = string("attn_out_85_cast_fp16")];
            tensor<int32, [4]> var_7726_perm_0 = const()[name = string("op_7726_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7728 = const()[name = string("op_7728"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_7726_cast_fp16 = transpose(perm = var_7726_perm_0, x = attn_out_85_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 1024]> input_173_cast_fp16 = reshape(shape = var_7728, x = var_7726_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_21_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_21_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736225792)))];
            tensor<fp16, [1, 16, 1024]> linear_150_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_self_attn_o_proj_weight_to_fp16, x = input_173_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_343_cast_fp16 = add(x = hidden_states_335_cast_fp16, y = linear_150_cast_fp16)[name = string("hidden_states_343_cast_fp16")];
            fp16 var_7738_promoted_to_fp16 = const()[name = string("op_7738_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7744_cast_fp16 = pow(x = hidden_states_343_cast_fp16, y = var_7738_promoted_to_fp16)[name = string("op_7744_cast_fp16")];
            tensor<int32, [1]> variance_87_axes_0 = const()[name = string("variance_87_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_87_keep_dims_0 = const()[name = string("variance_87_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_87_cast_fp16 = reduce_mean(axes = variance_87_axes_0, keep_dims = variance_87_keep_dims_0, x = var_7744_cast_fp16)[name = string("variance_87_cast_fp16")];
            fp16 var_7747_to_fp16 = const()[name = string("op_7747_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7748_cast_fp16 = add(x = variance_87_cast_fp16, y = var_7747_to_fp16)[name = string("op_7748_cast_fp16")];
            fp32 var_7749_epsilon_0 = const()[name = string("op_7749_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7749_cast_fp16 = rsqrt(epsilon = var_7749_epsilon_0, x = var_7748_cast_fp16)[name = string("op_7749_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_347_cast_fp16 = mul(x = hidden_states_343_cast_fp16, y = var_7749_cast_fp16)[name = string("hidden_states_347_cast_fp16")];
            tensor<fp16, [1024]> layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738323008)))];
            tensor<fp16, [1, 16, 1024]> input_175_cast_fp16 = mul(x = layers_21_post_attention_layernorm_weight_to_fp16, y = hidden_states_347_cast_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_21_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_21_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738325120)))];
            tensor<fp16, [1, 16, 4096]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_21_mlp_gate_proj_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_7762_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_7762_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_21_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_21_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746713792)))];
            tensor<fp16, [1, 16, 4096]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_21_mlp_up_proj_weight_to_fp16, x = input_175_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_179_cast_fp16 = mul(x = var_7762_cast_fp16, y = linear_152_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_21_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_21_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(755102464)))];
            tensor<fp16, [1, 16, 1024]> linear_153_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_21_mlp_down_proj_weight_to_fp16, x = input_179_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_351_cast_fp16 = add(x = hidden_states_343_cast_fp16, y = linear_153_cast_fp16)[name = string("hidden_states_351_cast_fp16")];
            tensor<int32, [4]> lkc_89_begin_0 = const()[name = string("lkc_89_begin_0"), val = tensor<int32, [4]>([0, 22528, 0, 0])];
            tensor<int32, [4]> lkc_89_end_0 = const()[name = string("lkc_89_end_0"), val = tensor<int32, [4]>([1, 23552, 1, 1024])];
            tensor<bool, [4]> lkc_89_end_mask_0 = const()[name = string("lkc_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_89_cast_fp16 = slice_by_index(begin = lkc_89_begin_0, end = lkc_89_end_0, end_mask = lkc_89_end_mask_0, x = key_cache)[name = string("lkc_89_cast_fp16")];
            tensor<int32, [4]> lvc_89_begin_0 = const()[name = string("lvc_89_begin_0"), val = tensor<int32, [4]>([0, 22528, 0, 0])];
            tensor<int32, [4]> lvc_89_end_0 = const()[name = string("lvc_89_end_0"), val = tensor<int32, [4]>([1, 23552, 1, 1024])];
            tensor<bool, [4]> lvc_89_end_mask_0 = const()[name = string("lvc_89_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_89_cast_fp16 = slice_by_index(begin = lvc_89_begin_0, end = lvc_89_end_0, end_mask = lvc_89_end_mask_0, x = value_cache)[name = string("lvc_89_cast_fp16")];
            fp16 var_7793_promoted_to_fp16 = const()[name = string("op_7793_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7799_cast_fp16 = pow(x = hidden_states_351_cast_fp16, y = var_7793_promoted_to_fp16)[name = string("op_7799_cast_fp16")];
            tensor<int32, [1]> variance_89_axes_0 = const()[name = string("variance_89_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_89_keep_dims_0 = const()[name = string("variance_89_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_89_cast_fp16 = reduce_mean(axes = variance_89_axes_0, keep_dims = variance_89_keep_dims_0, x = var_7799_cast_fp16)[name = string("variance_89_cast_fp16")];
            fp16 var_7802_to_fp16 = const()[name = string("op_7802_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_7803_cast_fp16 = add(x = variance_89_cast_fp16, y = var_7802_to_fp16)[name = string("op_7803_cast_fp16")];
            fp32 var_7804_epsilon_0 = const()[name = string("op_7804_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7804_cast_fp16 = rsqrt(epsilon = var_7804_epsilon_0, x = var_7803_cast_fp16)[name = string("op_7804_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_355_cast_fp16 = mul(x = hidden_states_351_cast_fp16, y = var_7804_cast_fp16)[name = string("hidden_states_355_cast_fp16")];
            tensor<fp16, [1024]> layers_22_input_layernorm_weight_to_fp16 = const()[name = string("layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763491136)))];
            tensor<fp16, [1, 16, 1024]> hidden_89_cast_fp16 = mul(x = layers_22_input_layernorm_weight_to_fp16, y = hidden_states_355_cast_fp16)[name = string("hidden_89_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(763493248)))];
            tensor<fp16, [1, 16, 1024]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_q_proj_weight_to_fp16, x = hidden_89_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_7828 = const()[name = string("op_7828"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_133_cast_fp16 = reshape(shape = var_7828, x = linear_154_cast_fp16)[name = string("q_133_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(765590464)))];
            tensor<fp16, [1, 16, 1024]> linear_155_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_k_proj_weight_to_fp16, x = hidden_89_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_7835 = const()[name = string("op_7835"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_89_cast_fp16 = reshape(shape = var_7835, x = linear_155_cast_fp16)[name = string("k_89_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(767687680)))];
            tensor<fp16, [1, 16, 1024]> linear_156_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_v_proj_weight_to_fp16, x = hidden_89_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_7842 = const()[name = string("op_7842"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_45_cast_fp16 = reshape(shape = var_7842, x = linear_156_cast_fp16)[name = string("v_45_cast_fp16")];
            tensor<int32, [4]> q_135_perm_0 = const()[name = string("q_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_91_perm_0 = const()[name = string("k_91_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_22_perm_0 = const()[name = string("transpose_22_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_135_cast_fp16 = transpose(perm = q_135_perm_0, x = q_133_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 64]> var_7855_cast_fp16 = mul(x = q_135_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7855_cast_fp16")];
            tensor<int32, [4]> x1_89_begin_0 = const()[name = string("x1_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_89_end_0 = const()[name = string("x1_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_89_end_mask_0 = const()[name = string("x1_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_89_cast_fp16 = slice_by_index(begin = x1_89_begin_0, end = x1_89_end_0, end_mask = x1_89_end_mask_0, x = q_135_cast_fp16)[name = string("x1_89_cast_fp16")];
            tensor<int32, [4]> x2_89_begin_0 = const()[name = string("x2_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_89_end_0 = const()[name = string("x2_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_89_end_mask_0 = const()[name = string("x2_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_89_cast_fp16 = slice_by_index(begin = x2_89_begin_0, end = x2_89_end_0, end_mask = x2_89_end_mask_0, x = q_135_cast_fp16)[name = string("x2_89_cast_fp16")];
            fp16 const_271_promoted_to_fp16 = const()[name = string("const_271_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7876_cast_fp16 = mul(x = x2_89_cast_fp16, y = const_271_promoted_to_fp16)[name = string("op_7876_cast_fp16")];
            int32 var_7878 = const()[name = string("op_7878"), val = int32(-1)];
            bool var_7879_interleave_0 = const()[name = string("op_7879_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7879_cast_fp16 = concat(axis = var_7878, interleave = var_7879_interleave_0, values = (var_7876_cast_fp16, x1_89_cast_fp16))[name = string("op_7879_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7882_cast_fp16 = mul(x = var_7879_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7882_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_137_cast_fp16 = add(x = var_7855_cast_fp16, y = var_7882_cast_fp16)[name = string("q_137_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_91_cast_fp16 = transpose(perm = k_91_perm_0, x = k_89_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 16, 16, 64]> var_7887_cast_fp16 = mul(x = k_91_cast_fp16, y = var_1078_cast_fp16)[name = string("op_7887_cast_fp16")];
            tensor<int32, [4]> x1_91_begin_0 = const()[name = string("x1_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_91_end_0 = const()[name = string("x1_91_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_91_end_mask_0 = const()[name = string("x1_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_91_cast_fp16 = slice_by_index(begin = x1_91_begin_0, end = x1_91_end_0, end_mask = x1_91_end_mask_0, x = k_91_cast_fp16)[name = string("x1_91_cast_fp16")];
            tensor<int32, [4]> x2_91_begin_0 = const()[name = string("x2_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_91_end_0 = const()[name = string("x2_91_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_91_end_mask_0 = const()[name = string("x2_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_91_cast_fp16 = slice_by_index(begin = x2_91_begin_0, end = x2_91_end_0, end_mask = x2_91_end_mask_0, x = k_91_cast_fp16)[name = string("x2_91_cast_fp16")];
            fp16 const_274_promoted_to_fp16 = const()[name = string("const_274_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_7908_cast_fp16 = mul(x = x2_91_cast_fp16, y = const_274_promoted_to_fp16)[name = string("op_7908_cast_fp16")];
            int32 var_7910 = const()[name = string("op_7910"), val = int32(-1)];
            bool var_7911_interleave_0 = const()[name = string("op_7911_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_7911_cast_fp16 = concat(axis = var_7910, interleave = var_7911_interleave_0, values = (var_7908_cast_fp16, x1_91_cast_fp16))[name = string("op_7911_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7914_cast_fp16 = mul(x = var_7911_cast_fp16, y = var_1105_cast_fp16)[name = string("op_7914_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_89_cast_fp16 = add(x = var_7887_cast_fp16, y = var_7914_cast_fp16)[name = string("x_89_cast_fp16")];
            tensor<int32, [4]> var_7921 = const()[name = string("op_7921"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_7933 = const()[name = string("op_7933"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_7922_cast_fp16 = transpose(perm = var_7921, x = x_89_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 1024, 1, 16]> var_7934_cast_fp16 = reshape(shape = var_7933, x = var_7922_cast_fp16)[name = string("op_7934_cast_fp16")];
            tensor<int32, [1]> var_7936_axes_0 = const()[name = string("op_7936_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7936_cast_fp16 = squeeze(axes = var_7936_axes_0, x = var_7934_cast_fp16)[name = string("op_7936_cast_fp16")];
            bool var_7937_transpose_x_0 = const()[name = string("op_7937_transpose_x_0"), val = bool(false)];
            bool var_7937_transpose_y_0 = const()[name = string("op_7937_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7937_cast_fp16 = matmul(transpose_x = var_7937_transpose_x_0, transpose_y = var_7937_transpose_y_0, x = var_7936_cast_fp16, y = block_cache_map)[name = string("op_7937_cast_fp16")];
            tensor<int32, [1]> nk_full_45_axes_0 = const()[name = string("nk_full_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_45_cast_fp16 = expand_dims(axes = nk_full_45_axes_0, x = var_7937_cast_fp16)[name = string("nk_full_45_cast_fp16")];
            tensor<int32, [4]> var_7956 = const()[name = string("op_7956"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_22_cast_fp16 = transpose(perm = transpose_22_perm_0, x = v_45_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 1024, 1, 16]> var_7957_cast_fp16 = reshape(shape = var_7956, x = transpose_22_cast_fp16)[name = string("op_7957_cast_fp16")];
            tensor<int32, [1]> var_7959_axes_0 = const()[name = string("op_7959_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_7959_cast_fp16 = squeeze(axes = var_7959_axes_0, x = var_7957_cast_fp16)[name = string("op_7959_cast_fp16")];
            bool var_7960_transpose_x_0 = const()[name = string("op_7960_transpose_x_0"), val = bool(false)];
            bool var_7960_transpose_y_0 = const()[name = string("op_7960_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_7960_cast_fp16 = matmul(transpose_x = var_7960_transpose_x_0, transpose_y = var_7960_transpose_y_0, x = var_7959_cast_fp16, y = block_cache_map)[name = string("op_7960_cast_fp16")];
            tensor<int32, [1]> nv_full_45_axes_0 = const()[name = string("nv_full_45_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_45_cast_fp16 = expand_dims(axes = nv_full_45_axes_0, x = var_7960_cast_fp16)[name = string("nv_full_45_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7966_cast_fp16 = mul(x = lkc_89_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7966_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7967_cast_fp16 = mul(x = nk_full_45_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7967_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_91_cast_fp16 = add(x = var_7966_cast_fp16, y = var_7967_cast_fp16)[name = string("lkc_91_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7973_cast_fp16 = mul(x = lvc_89_cast_fp16, y = var_1189_cast_fp16)[name = string("op_7973_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_7974_cast_fp16 = mul(x = nv_full_45_cast_fp16, y = update_mask_cast_fp16)[name = string("op_7974_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_91_cast_fp16 = add(x = var_7973_cast_fp16, y = var_7974_cast_fp16)[name = string("lvc_91_cast_fp16")];
            tensor<int32, [1]> var_7978_axes_0 = const()[name = string("op_7978_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7978_cast_fp16 = squeeze(axes = var_7978_axes_0, x = lkc_91_cast_fp16)[name = string("op_7978_cast_fp16")];
            tensor<int32, [4]> var_7982 = const()[name = string("op_7982"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_45_cast_fp16 = reshape(shape = var_7982, x = var_7978_cast_fp16)[name = string("kc_45_cast_fp16")];
            tensor<int32, [1]> var_7985_axes_0 = const()[name = string("op_7985_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_7985_cast_fp16 = squeeze(axes = var_7985_axes_0, x = lvc_91_cast_fp16)[name = string("op_7985_cast_fp16")];
            tensor<int32, [4]> var_7989 = const()[name = string("op_7989"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_45_cast_fp16 = reshape(shape = var_7989, x = var_7985_cast_fp16)[name = string("vc_45_cast_fp16")];
            bool var_7991_transpose_x_0 = const()[name = string("op_7991_transpose_x_0"), val = bool(false)];
            bool var_7991_transpose_y_0 = const()[name = string("op_7991_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_7991_cast_fp16 = matmul(transpose_x = var_7991_transpose_x_0, transpose_y = var_7991_transpose_y_0, x = q_137_cast_fp16, y = kc_45_cast_fp16)[name = string("op_7991_cast_fp16")];
            fp16 _inversed_scores_133_y_0_to_fp16 = const()[name = string("_inversed_scores_133_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_133_cast_fp16 = mul(x = var_7991_cast_fp16, y = _inversed_scores_133_y_0_to_fp16)[name = string("_inversed_scores_133_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_135_cast_fp16 = add(x = _inversed_scores_133_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_135_cast_fp16")];
            int32 var_8020 = const()[name = string("op_8020"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_8022_cast_fp16 = softmax(axis = var_8020, x = scores_135_cast_fp16)[name = string("op_8022_cast_fp16")];
            bool attn_out_89_transpose_x_1 = const()[name = string("attn_out_89_transpose_x_1"), val = bool(false)];
            bool attn_out_89_transpose_y_1 = const()[name = string("attn_out_89_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_89_cast_fp16 = matmul(transpose_x = attn_out_89_transpose_x_1, transpose_y = attn_out_89_transpose_y_1, x = var_8022_cast_fp16, y = vc_45_cast_fp16)[name = string("attn_out_89_cast_fp16")];
            tensor<int32, [4]> var_8034_perm_0 = const()[name = string("op_8034_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8036 = const()[name = string("op_8036"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_8034_cast_fp16 = transpose(perm = var_8034_perm_0, x = attn_out_89_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 1024]> input_181_cast_fp16 = reshape(shape = var_8036, x = var_8034_cast_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_22_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_22_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(769784896)))];
            tensor<fp16, [1, 16, 1024]> linear_157_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_self_attn_o_proj_weight_to_fp16, x = input_181_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_359_cast_fp16 = add(x = hidden_states_351_cast_fp16, y = linear_157_cast_fp16)[name = string("hidden_states_359_cast_fp16")];
            fp16 var_8046_promoted_to_fp16 = const()[name = string("op_8046_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8052_cast_fp16 = pow(x = hidden_states_359_cast_fp16, y = var_8046_promoted_to_fp16)[name = string("op_8052_cast_fp16")];
            tensor<int32, [1]> variance_91_axes_0 = const()[name = string("variance_91_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_91_keep_dims_0 = const()[name = string("variance_91_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_91_cast_fp16 = reduce_mean(axes = variance_91_axes_0, keep_dims = variance_91_keep_dims_0, x = var_8052_cast_fp16)[name = string("variance_91_cast_fp16")];
            fp16 var_8055_to_fp16 = const()[name = string("op_8055_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8056_cast_fp16 = add(x = variance_91_cast_fp16, y = var_8055_to_fp16)[name = string("op_8056_cast_fp16")];
            fp32 var_8057_epsilon_0 = const()[name = string("op_8057_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8057_cast_fp16 = rsqrt(epsilon = var_8057_epsilon_0, x = var_8056_cast_fp16)[name = string("op_8057_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_363_cast_fp16 = mul(x = hidden_states_359_cast_fp16, y = var_8057_cast_fp16)[name = string("hidden_states_363_cast_fp16")];
            tensor<fp16, [1024]> layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771882112)))];
            tensor<fp16, [1, 16, 1024]> input_183_cast_fp16 = mul(x = layers_22_post_attention_layernorm_weight_to_fp16, y = hidden_states_363_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_22_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_22_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(771884224)))];
            tensor<fp16, [1, 16, 4096]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_22_mlp_gate_proj_weight_to_fp16, x = input_183_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_8070_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8070_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_22_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_22_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(780272896)))];
            tensor<fp16, [1, 16, 4096]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_22_mlp_up_proj_weight_to_fp16, x = input_183_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_187_cast_fp16 = mul(x = var_8070_cast_fp16, y = linear_159_cast_fp16)[name = string("input_187_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_22_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_22_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(788661568)))];
            tensor<fp16, [1, 16, 1024]> linear_160_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_22_mlp_down_proj_weight_to_fp16, x = input_187_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_367_cast_fp16 = add(x = hidden_states_359_cast_fp16, y = linear_160_cast_fp16)[name = string("hidden_states_367_cast_fp16")];
            tensor<int32, [4]> lkc_93_begin_0 = const()[name = string("lkc_93_begin_0"), val = tensor<int32, [4]>([0, 23552, 0, 0])];
            tensor<int32, [4]> lkc_93_end_0 = const()[name = string("lkc_93_end_0"), val = tensor<int32, [4]>([1, 24576, 1, 1024])];
            tensor<bool, [4]> lkc_93_end_mask_0 = const()[name = string("lkc_93_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_93_cast_fp16 = slice_by_index(begin = lkc_93_begin_0, end = lkc_93_end_0, end_mask = lkc_93_end_mask_0, x = key_cache)[name = string("lkc_93_cast_fp16")];
            tensor<int32, [4]> lvc_93_begin_0 = const()[name = string("lvc_93_begin_0"), val = tensor<int32, [4]>([0, 23552, 0, 0])];
            tensor<int32, [4]> lvc_93_end_0 = const()[name = string("lvc_93_end_0"), val = tensor<int32, [4]>([1, 24576, 1, 1024])];
            tensor<bool, [4]> lvc_93_end_mask_0 = const()[name = string("lvc_93_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_93_cast_fp16 = slice_by_index(begin = lvc_93_begin_0, end = lvc_93_end_0, end_mask = lvc_93_end_mask_0, x = value_cache)[name = string("lvc_93_cast_fp16")];
            fp16 var_8101_promoted_to_fp16 = const()[name = string("op_8101_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8107_cast_fp16 = pow(x = hidden_states_367_cast_fp16, y = var_8101_promoted_to_fp16)[name = string("op_8107_cast_fp16")];
            tensor<int32, [1]> variance_93_axes_0 = const()[name = string("variance_93_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_93_keep_dims_0 = const()[name = string("variance_93_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_93_cast_fp16 = reduce_mean(axes = variance_93_axes_0, keep_dims = variance_93_keep_dims_0, x = var_8107_cast_fp16)[name = string("variance_93_cast_fp16")];
            fp16 var_8110_to_fp16 = const()[name = string("op_8110_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8111_cast_fp16 = add(x = variance_93_cast_fp16, y = var_8110_to_fp16)[name = string("op_8111_cast_fp16")];
            fp32 var_8112_epsilon_0 = const()[name = string("op_8112_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8112_cast_fp16 = rsqrt(epsilon = var_8112_epsilon_0, x = var_8111_cast_fp16)[name = string("op_8112_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_371_cast_fp16 = mul(x = hidden_states_367_cast_fp16, y = var_8112_cast_fp16)[name = string("hidden_states_371_cast_fp16")];
            tensor<fp16, [1024]> layers_23_input_layernorm_weight_to_fp16 = const()[name = string("layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(797050240)))];
            tensor<fp16, [1, 16, 1024]> hidden_93_cast_fp16 = mul(x = layers_23_input_layernorm_weight_to_fp16, y = hidden_states_371_cast_fp16)[name = string("hidden_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(797052352)))];
            tensor<fp16, [1, 16, 1024]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_q_proj_weight_to_fp16, x = hidden_93_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_8136 = const()[name = string("op_8136"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_139_cast_fp16 = reshape(shape = var_8136, x = linear_161_cast_fp16)[name = string("q_139_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(799149568)))];
            tensor<fp16, [1, 16, 1024]> linear_162_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_k_proj_weight_to_fp16, x = hidden_93_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_8143 = const()[name = string("op_8143"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_93_cast_fp16 = reshape(shape = var_8143, x = linear_162_cast_fp16)[name = string("k_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(801246784)))];
            tensor<fp16, [1, 16, 1024]> linear_163_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_v_proj_weight_to_fp16, x = hidden_93_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8150 = const()[name = string("op_8150"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_47_cast_fp16 = reshape(shape = var_8150, x = linear_163_cast_fp16)[name = string("v_47_cast_fp16")];
            tensor<int32, [4]> q_141_perm_0 = const()[name = string("q_141_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_95_perm_0 = const()[name = string("k_95_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_23_perm_0 = const()[name = string("transpose_23_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_141_cast_fp16 = transpose(perm = q_141_perm_0, x = q_139_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 64]> var_8163_cast_fp16 = mul(x = q_141_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8163_cast_fp16")];
            tensor<int32, [4]> x1_93_begin_0 = const()[name = string("x1_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_93_end_0 = const()[name = string("x1_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_93_end_mask_0 = const()[name = string("x1_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_93_cast_fp16 = slice_by_index(begin = x1_93_begin_0, end = x1_93_end_0, end_mask = x1_93_end_mask_0, x = q_141_cast_fp16)[name = string("x1_93_cast_fp16")];
            tensor<int32, [4]> x2_93_begin_0 = const()[name = string("x2_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_93_end_0 = const()[name = string("x2_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_93_end_mask_0 = const()[name = string("x2_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_93_cast_fp16 = slice_by_index(begin = x2_93_begin_0, end = x2_93_end_0, end_mask = x2_93_end_mask_0, x = q_141_cast_fp16)[name = string("x2_93_cast_fp16")];
            fp16 const_283_promoted_to_fp16 = const()[name = string("const_283_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8184_cast_fp16 = mul(x = x2_93_cast_fp16, y = const_283_promoted_to_fp16)[name = string("op_8184_cast_fp16")];
            int32 var_8186 = const()[name = string("op_8186"), val = int32(-1)];
            bool var_8187_interleave_0 = const()[name = string("op_8187_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8187_cast_fp16 = concat(axis = var_8186, interleave = var_8187_interleave_0, values = (var_8184_cast_fp16, x1_93_cast_fp16))[name = string("op_8187_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8190_cast_fp16 = mul(x = var_8187_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8190_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_143_cast_fp16 = add(x = var_8163_cast_fp16, y = var_8190_cast_fp16)[name = string("q_143_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_95_cast_fp16 = transpose(perm = k_95_perm_0, x = k_93_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 16, 16, 64]> var_8195_cast_fp16 = mul(x = k_95_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8195_cast_fp16")];
            tensor<int32, [4]> x1_95_begin_0 = const()[name = string("x1_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_95_end_0 = const()[name = string("x1_95_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_95_end_mask_0 = const()[name = string("x1_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_95_cast_fp16 = slice_by_index(begin = x1_95_begin_0, end = x1_95_end_0, end_mask = x1_95_end_mask_0, x = k_95_cast_fp16)[name = string("x1_95_cast_fp16")];
            tensor<int32, [4]> x2_95_begin_0 = const()[name = string("x2_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_95_end_0 = const()[name = string("x2_95_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_95_end_mask_0 = const()[name = string("x2_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_95_cast_fp16 = slice_by_index(begin = x2_95_begin_0, end = x2_95_end_0, end_mask = x2_95_end_mask_0, x = k_95_cast_fp16)[name = string("x2_95_cast_fp16")];
            fp16 const_286_promoted_to_fp16 = const()[name = string("const_286_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8216_cast_fp16 = mul(x = x2_95_cast_fp16, y = const_286_promoted_to_fp16)[name = string("op_8216_cast_fp16")];
            int32 var_8218 = const()[name = string("op_8218"), val = int32(-1)];
            bool var_8219_interleave_0 = const()[name = string("op_8219_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8219_cast_fp16 = concat(axis = var_8218, interleave = var_8219_interleave_0, values = (var_8216_cast_fp16, x1_95_cast_fp16))[name = string("op_8219_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8222_cast_fp16 = mul(x = var_8219_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8222_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_93_cast_fp16 = add(x = var_8195_cast_fp16, y = var_8222_cast_fp16)[name = string("x_93_cast_fp16")];
            tensor<int32, [4]> var_8229 = const()[name = string("op_8229"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8241 = const()[name = string("op_8241"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_8230_cast_fp16 = transpose(perm = var_8229, x = x_93_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 1024, 1, 16]> var_8242_cast_fp16 = reshape(shape = var_8241, x = var_8230_cast_fp16)[name = string("op_8242_cast_fp16")];
            tensor<int32, [1]> var_8244_axes_0 = const()[name = string("op_8244_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8244_cast_fp16 = squeeze(axes = var_8244_axes_0, x = var_8242_cast_fp16)[name = string("op_8244_cast_fp16")];
            bool var_8245_transpose_x_0 = const()[name = string("op_8245_transpose_x_0"), val = bool(false)];
            bool var_8245_transpose_y_0 = const()[name = string("op_8245_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8245_cast_fp16 = matmul(transpose_x = var_8245_transpose_x_0, transpose_y = var_8245_transpose_y_0, x = var_8244_cast_fp16, y = block_cache_map)[name = string("op_8245_cast_fp16")];
            tensor<int32, [1]> nk_full_47_axes_0 = const()[name = string("nk_full_47_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_47_cast_fp16 = expand_dims(axes = nk_full_47_axes_0, x = var_8245_cast_fp16)[name = string("nk_full_47_cast_fp16")];
            tensor<int32, [4]> var_8264 = const()[name = string("op_8264"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_23_cast_fp16 = transpose(perm = transpose_23_perm_0, x = v_47_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 1024, 1, 16]> var_8265_cast_fp16 = reshape(shape = var_8264, x = transpose_23_cast_fp16)[name = string("op_8265_cast_fp16")];
            tensor<int32, [1]> var_8267_axes_0 = const()[name = string("op_8267_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8267_cast_fp16 = squeeze(axes = var_8267_axes_0, x = var_8265_cast_fp16)[name = string("op_8267_cast_fp16")];
            bool var_8268_transpose_x_0 = const()[name = string("op_8268_transpose_x_0"), val = bool(false)];
            bool var_8268_transpose_y_0 = const()[name = string("op_8268_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8268_cast_fp16 = matmul(transpose_x = var_8268_transpose_x_0, transpose_y = var_8268_transpose_y_0, x = var_8267_cast_fp16, y = block_cache_map)[name = string("op_8268_cast_fp16")];
            tensor<int32, [1]> nv_full_47_axes_0 = const()[name = string("nv_full_47_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_47_cast_fp16 = expand_dims(axes = nv_full_47_axes_0, x = var_8268_cast_fp16)[name = string("nv_full_47_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8274_cast_fp16 = mul(x = lkc_93_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8274_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8275_cast_fp16 = mul(x = nk_full_47_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8275_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_95_cast_fp16 = add(x = var_8274_cast_fp16, y = var_8275_cast_fp16)[name = string("lkc_95_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8281_cast_fp16 = mul(x = lvc_93_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8281_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8282_cast_fp16 = mul(x = nv_full_47_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8282_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_95_cast_fp16 = add(x = var_8281_cast_fp16, y = var_8282_cast_fp16)[name = string("lvc_95_cast_fp16")];
            tensor<int32, [1]> var_8286_axes_0 = const()[name = string("op_8286_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8286_cast_fp16 = squeeze(axes = var_8286_axes_0, x = lkc_95_cast_fp16)[name = string("op_8286_cast_fp16")];
            tensor<int32, [4]> var_8290 = const()[name = string("op_8290"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_47_cast_fp16 = reshape(shape = var_8290, x = var_8286_cast_fp16)[name = string("kc_47_cast_fp16")];
            tensor<int32, [1]> var_8293_axes_0 = const()[name = string("op_8293_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8293_cast_fp16 = squeeze(axes = var_8293_axes_0, x = lvc_95_cast_fp16)[name = string("op_8293_cast_fp16")];
            tensor<int32, [4]> var_8297 = const()[name = string("op_8297"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_47_cast_fp16 = reshape(shape = var_8297, x = var_8293_cast_fp16)[name = string("vc_47_cast_fp16")];
            bool var_8299_transpose_x_0 = const()[name = string("op_8299_transpose_x_0"), val = bool(false)];
            bool var_8299_transpose_y_0 = const()[name = string("op_8299_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_8299_cast_fp16 = matmul(transpose_x = var_8299_transpose_x_0, transpose_y = var_8299_transpose_y_0, x = q_143_cast_fp16, y = kc_47_cast_fp16)[name = string("op_8299_cast_fp16")];
            fp16 _inversed_scores_139_y_0_to_fp16 = const()[name = string("_inversed_scores_139_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_139_cast_fp16 = mul(x = var_8299_cast_fp16, y = _inversed_scores_139_y_0_to_fp16)[name = string("_inversed_scores_139_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_141_cast_fp16 = add(x = _inversed_scores_139_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_141_cast_fp16")];
            int32 var_8328 = const()[name = string("op_8328"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_8330_cast_fp16 = softmax(axis = var_8328, x = scores_141_cast_fp16)[name = string("op_8330_cast_fp16")];
            bool attn_out_93_transpose_x_1 = const()[name = string("attn_out_93_transpose_x_1"), val = bool(false)];
            bool attn_out_93_transpose_y_1 = const()[name = string("attn_out_93_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_93_cast_fp16 = matmul(transpose_x = attn_out_93_transpose_x_1, transpose_y = attn_out_93_transpose_y_1, x = var_8330_cast_fp16, y = vc_47_cast_fp16)[name = string("attn_out_93_cast_fp16")];
            tensor<int32, [4]> var_8342_perm_0 = const()[name = string("op_8342_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8344 = const()[name = string("op_8344"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_8342_cast_fp16 = transpose(perm = var_8342_perm_0, x = attn_out_93_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 1024]> input_189_cast_fp16 = reshape(shape = var_8344, x = var_8342_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_23_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_23_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(803344000)))];
            tensor<fp16, [1, 16, 1024]> linear_164_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_self_attn_o_proj_weight_to_fp16, x = input_189_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_375_cast_fp16 = add(x = hidden_states_367_cast_fp16, y = linear_164_cast_fp16)[name = string("hidden_states_375_cast_fp16")];
            fp16 var_8354_promoted_to_fp16 = const()[name = string("op_8354_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8360_cast_fp16 = pow(x = hidden_states_375_cast_fp16, y = var_8354_promoted_to_fp16)[name = string("op_8360_cast_fp16")];
            tensor<int32, [1]> variance_95_axes_0 = const()[name = string("variance_95_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_95_keep_dims_0 = const()[name = string("variance_95_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_95_cast_fp16 = reduce_mean(axes = variance_95_axes_0, keep_dims = variance_95_keep_dims_0, x = var_8360_cast_fp16)[name = string("variance_95_cast_fp16")];
            fp16 var_8363_to_fp16 = const()[name = string("op_8363_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8364_cast_fp16 = add(x = variance_95_cast_fp16, y = var_8363_to_fp16)[name = string("op_8364_cast_fp16")];
            fp32 var_8365_epsilon_0 = const()[name = string("op_8365_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8365_cast_fp16 = rsqrt(epsilon = var_8365_epsilon_0, x = var_8364_cast_fp16)[name = string("op_8365_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_379_cast_fp16 = mul(x = hidden_states_375_cast_fp16, y = var_8365_cast_fp16)[name = string("hidden_states_379_cast_fp16")];
            tensor<fp16, [1024]> layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(805441216)))];
            tensor<fp16, [1, 16, 1024]> input_191_cast_fp16 = mul(x = layers_23_post_attention_layernorm_weight_to_fp16, y = hidden_states_379_cast_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_23_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_23_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(805443328)))];
            tensor<fp16, [1, 16, 4096]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_23_mlp_gate_proj_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_8378_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_8378_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_23_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_23_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(813832000)))];
            tensor<fp16, [1, 16, 4096]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_23_mlp_up_proj_weight_to_fp16, x = input_191_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_195_cast_fp16 = mul(x = var_8378_cast_fp16, y = linear_166_cast_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_23_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_23_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(822220672)))];
            tensor<fp16, [1, 16, 1024]> linear_167_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_23_mlp_down_proj_weight_to_fp16, x = input_195_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_383_cast_fp16 = add(x = hidden_states_375_cast_fp16, y = linear_167_cast_fp16)[name = string("hidden_states_383_cast_fp16")];
            tensor<int32, [4]> lkc_97_begin_0 = const()[name = string("lkc_97_begin_0"), val = tensor<int32, [4]>([0, 24576, 0, 0])];
            tensor<int32, [4]> lkc_97_end_0 = const()[name = string("lkc_97_end_0"), val = tensor<int32, [4]>([1, 25600, 1, 1024])];
            tensor<bool, [4]> lkc_97_end_mask_0 = const()[name = string("lkc_97_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_97_cast_fp16 = slice_by_index(begin = lkc_97_begin_0, end = lkc_97_end_0, end_mask = lkc_97_end_mask_0, x = key_cache)[name = string("lkc_97_cast_fp16")];
            tensor<int32, [4]> lvc_97_begin_0 = const()[name = string("lvc_97_begin_0"), val = tensor<int32, [4]>([0, 24576, 0, 0])];
            tensor<int32, [4]> lvc_97_end_0 = const()[name = string("lvc_97_end_0"), val = tensor<int32, [4]>([1, 25600, 1, 1024])];
            tensor<bool, [4]> lvc_97_end_mask_0 = const()[name = string("lvc_97_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_97_cast_fp16 = slice_by_index(begin = lvc_97_begin_0, end = lvc_97_end_0, end_mask = lvc_97_end_mask_0, x = value_cache)[name = string("lvc_97_cast_fp16")];
            fp16 var_8409_promoted_to_fp16 = const()[name = string("op_8409_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8415_cast_fp16 = pow(x = hidden_states_383_cast_fp16, y = var_8409_promoted_to_fp16)[name = string("op_8415_cast_fp16")];
            tensor<int32, [1]> variance_97_axes_0 = const()[name = string("variance_97_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_97_keep_dims_0 = const()[name = string("variance_97_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_97_cast_fp16 = reduce_mean(axes = variance_97_axes_0, keep_dims = variance_97_keep_dims_0, x = var_8415_cast_fp16)[name = string("variance_97_cast_fp16")];
            fp16 var_8418_to_fp16 = const()[name = string("op_8418_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8419_cast_fp16 = add(x = variance_97_cast_fp16, y = var_8418_to_fp16)[name = string("op_8419_cast_fp16")];
            fp32 var_8420_epsilon_0 = const()[name = string("op_8420_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8420_cast_fp16 = rsqrt(epsilon = var_8420_epsilon_0, x = var_8419_cast_fp16)[name = string("op_8420_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_387_cast_fp16 = mul(x = hidden_states_383_cast_fp16, y = var_8420_cast_fp16)[name = string("hidden_states_387_cast_fp16")];
            tensor<fp16, [1024]> layers_24_input_layernorm_weight_to_fp16 = const()[name = string("layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(830609344)))];
            tensor<fp16, [1, 16, 1024]> hidden_97_cast_fp16 = mul(x = layers_24_input_layernorm_weight_to_fp16, y = hidden_states_387_cast_fp16)[name = string("hidden_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(830611456)))];
            tensor<fp16, [1, 16, 1024]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_q_proj_weight_to_fp16, x = hidden_97_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [4]> var_8444 = const()[name = string("op_8444"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_145_cast_fp16 = reshape(shape = var_8444, x = linear_168_cast_fp16)[name = string("q_145_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(832708672)))];
            tensor<fp16, [1, 16, 1024]> linear_169_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_k_proj_weight_to_fp16, x = hidden_97_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [4]> var_8451 = const()[name = string("op_8451"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_97_cast_fp16 = reshape(shape = var_8451, x = linear_169_cast_fp16)[name = string("k_97_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(834805888)))];
            tensor<fp16, [1, 16, 1024]> linear_170_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_v_proj_weight_to_fp16, x = hidden_97_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> var_8458 = const()[name = string("op_8458"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_49_cast_fp16 = reshape(shape = var_8458, x = linear_170_cast_fp16)[name = string("v_49_cast_fp16")];
            tensor<int32, [4]> q_147_perm_0 = const()[name = string("q_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_99_perm_0 = const()[name = string("k_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_24_perm_0 = const()[name = string("transpose_24_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_147_cast_fp16 = transpose(perm = q_147_perm_0, x = q_145_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 64]> var_8471_cast_fp16 = mul(x = q_147_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8471_cast_fp16")];
            tensor<int32, [4]> x1_97_begin_0 = const()[name = string("x1_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_97_end_0 = const()[name = string("x1_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_97_end_mask_0 = const()[name = string("x1_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_97_cast_fp16 = slice_by_index(begin = x1_97_begin_0, end = x1_97_end_0, end_mask = x1_97_end_mask_0, x = q_147_cast_fp16)[name = string("x1_97_cast_fp16")];
            tensor<int32, [4]> x2_97_begin_0 = const()[name = string("x2_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_97_end_0 = const()[name = string("x2_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_97_end_mask_0 = const()[name = string("x2_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_97_cast_fp16 = slice_by_index(begin = x2_97_begin_0, end = x2_97_end_0, end_mask = x2_97_end_mask_0, x = q_147_cast_fp16)[name = string("x2_97_cast_fp16")];
            fp16 const_295_promoted_to_fp16 = const()[name = string("const_295_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8492_cast_fp16 = mul(x = x2_97_cast_fp16, y = const_295_promoted_to_fp16)[name = string("op_8492_cast_fp16")];
            int32 var_8494 = const()[name = string("op_8494"), val = int32(-1)];
            bool var_8495_interleave_0 = const()[name = string("op_8495_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8495_cast_fp16 = concat(axis = var_8494, interleave = var_8495_interleave_0, values = (var_8492_cast_fp16, x1_97_cast_fp16))[name = string("op_8495_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8498_cast_fp16 = mul(x = var_8495_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8498_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_149_cast_fp16 = add(x = var_8471_cast_fp16, y = var_8498_cast_fp16)[name = string("q_149_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_99_cast_fp16 = transpose(perm = k_99_perm_0, x = k_97_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 16, 16, 64]> var_8503_cast_fp16 = mul(x = k_99_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8503_cast_fp16")];
            tensor<int32, [4]> x1_99_begin_0 = const()[name = string("x1_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_99_end_0 = const()[name = string("x1_99_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_99_end_mask_0 = const()[name = string("x1_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_99_cast_fp16 = slice_by_index(begin = x1_99_begin_0, end = x1_99_end_0, end_mask = x1_99_end_mask_0, x = k_99_cast_fp16)[name = string("x1_99_cast_fp16")];
            tensor<int32, [4]> x2_99_begin_0 = const()[name = string("x2_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_99_end_0 = const()[name = string("x2_99_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_99_end_mask_0 = const()[name = string("x2_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_99_cast_fp16 = slice_by_index(begin = x2_99_begin_0, end = x2_99_end_0, end_mask = x2_99_end_mask_0, x = k_99_cast_fp16)[name = string("x2_99_cast_fp16")];
            fp16 const_298_promoted_to_fp16 = const()[name = string("const_298_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8524_cast_fp16 = mul(x = x2_99_cast_fp16, y = const_298_promoted_to_fp16)[name = string("op_8524_cast_fp16")];
            int32 var_8526 = const()[name = string("op_8526"), val = int32(-1)];
            bool var_8527_interleave_0 = const()[name = string("op_8527_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8527_cast_fp16 = concat(axis = var_8526, interleave = var_8527_interleave_0, values = (var_8524_cast_fp16, x1_99_cast_fp16))[name = string("op_8527_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8530_cast_fp16 = mul(x = var_8527_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8530_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_97_cast_fp16 = add(x = var_8503_cast_fp16, y = var_8530_cast_fp16)[name = string("x_97_cast_fp16")];
            tensor<int32, [4]> var_8537 = const()[name = string("op_8537"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8549 = const()[name = string("op_8549"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_8538_cast_fp16 = transpose(perm = var_8537, x = x_97_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 1024, 1, 16]> var_8550_cast_fp16 = reshape(shape = var_8549, x = var_8538_cast_fp16)[name = string("op_8550_cast_fp16")];
            tensor<int32, [1]> var_8552_axes_0 = const()[name = string("op_8552_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8552_cast_fp16 = squeeze(axes = var_8552_axes_0, x = var_8550_cast_fp16)[name = string("op_8552_cast_fp16")];
            bool var_8553_transpose_x_0 = const()[name = string("op_8553_transpose_x_0"), val = bool(false)];
            bool var_8553_transpose_y_0 = const()[name = string("op_8553_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8553_cast_fp16 = matmul(transpose_x = var_8553_transpose_x_0, transpose_y = var_8553_transpose_y_0, x = var_8552_cast_fp16, y = block_cache_map)[name = string("op_8553_cast_fp16")];
            tensor<int32, [1]> nk_full_49_axes_0 = const()[name = string("nk_full_49_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_49_cast_fp16 = expand_dims(axes = nk_full_49_axes_0, x = var_8553_cast_fp16)[name = string("nk_full_49_cast_fp16")];
            tensor<int32, [4]> var_8572 = const()[name = string("op_8572"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_24_cast_fp16 = transpose(perm = transpose_24_perm_0, x = v_49_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 1024, 1, 16]> var_8573_cast_fp16 = reshape(shape = var_8572, x = transpose_24_cast_fp16)[name = string("op_8573_cast_fp16")];
            tensor<int32, [1]> var_8575_axes_0 = const()[name = string("op_8575_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8575_cast_fp16 = squeeze(axes = var_8575_axes_0, x = var_8573_cast_fp16)[name = string("op_8575_cast_fp16")];
            bool var_8576_transpose_x_0 = const()[name = string("op_8576_transpose_x_0"), val = bool(false)];
            bool var_8576_transpose_y_0 = const()[name = string("op_8576_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8576_cast_fp16 = matmul(transpose_x = var_8576_transpose_x_0, transpose_y = var_8576_transpose_y_0, x = var_8575_cast_fp16, y = block_cache_map)[name = string("op_8576_cast_fp16")];
            tensor<int32, [1]> nv_full_49_axes_0 = const()[name = string("nv_full_49_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_49_cast_fp16 = expand_dims(axes = nv_full_49_axes_0, x = var_8576_cast_fp16)[name = string("nv_full_49_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8582_cast_fp16 = mul(x = lkc_97_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8582_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8583_cast_fp16 = mul(x = nk_full_49_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8583_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_99_cast_fp16 = add(x = var_8582_cast_fp16, y = var_8583_cast_fp16)[name = string("lkc_99_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8589_cast_fp16 = mul(x = lvc_97_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8589_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8590_cast_fp16 = mul(x = nv_full_49_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8590_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_99_cast_fp16 = add(x = var_8589_cast_fp16, y = var_8590_cast_fp16)[name = string("lvc_99_cast_fp16")];
            tensor<int32, [1]> var_8594_axes_0 = const()[name = string("op_8594_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8594_cast_fp16 = squeeze(axes = var_8594_axes_0, x = lkc_99_cast_fp16)[name = string("op_8594_cast_fp16")];
            tensor<int32, [4]> var_8598 = const()[name = string("op_8598"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_49_cast_fp16 = reshape(shape = var_8598, x = var_8594_cast_fp16)[name = string("kc_49_cast_fp16")];
            tensor<int32, [1]> var_8601_axes_0 = const()[name = string("op_8601_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8601_cast_fp16 = squeeze(axes = var_8601_axes_0, x = lvc_99_cast_fp16)[name = string("op_8601_cast_fp16")];
            tensor<int32, [4]> var_8605 = const()[name = string("op_8605"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_49_cast_fp16 = reshape(shape = var_8605, x = var_8601_cast_fp16)[name = string("vc_49_cast_fp16")];
            bool var_8607_transpose_x_0 = const()[name = string("op_8607_transpose_x_0"), val = bool(false)];
            bool var_8607_transpose_y_0 = const()[name = string("op_8607_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_8607_cast_fp16 = matmul(transpose_x = var_8607_transpose_x_0, transpose_y = var_8607_transpose_y_0, x = q_149_cast_fp16, y = kc_49_cast_fp16)[name = string("op_8607_cast_fp16")];
            fp16 _inversed_scores_145_y_0_to_fp16 = const()[name = string("_inversed_scores_145_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_145_cast_fp16 = mul(x = var_8607_cast_fp16, y = _inversed_scores_145_y_0_to_fp16)[name = string("_inversed_scores_145_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_147_cast_fp16 = add(x = _inversed_scores_145_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_147_cast_fp16")];
            int32 var_8636 = const()[name = string("op_8636"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_8638_cast_fp16 = softmax(axis = var_8636, x = scores_147_cast_fp16)[name = string("op_8638_cast_fp16")];
            bool attn_out_97_transpose_x_1 = const()[name = string("attn_out_97_transpose_x_1"), val = bool(false)];
            bool attn_out_97_transpose_y_1 = const()[name = string("attn_out_97_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_97_cast_fp16 = matmul(transpose_x = attn_out_97_transpose_x_1, transpose_y = attn_out_97_transpose_y_1, x = var_8638_cast_fp16, y = vc_49_cast_fp16)[name = string("attn_out_97_cast_fp16")];
            tensor<int32, [4]> var_8650_perm_0 = const()[name = string("op_8650_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8652 = const()[name = string("op_8652"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_8650_cast_fp16 = transpose(perm = var_8650_perm_0, x = attn_out_97_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 1024]> input_197_cast_fp16 = reshape(shape = var_8652, x = var_8650_cast_fp16)[name = string("input_197_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_24_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_24_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(836903104)))];
            tensor<fp16, [1, 16, 1024]> linear_171_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_self_attn_o_proj_weight_to_fp16, x = input_197_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_391_cast_fp16 = add(x = hidden_states_383_cast_fp16, y = linear_171_cast_fp16)[name = string("hidden_states_391_cast_fp16")];
            fp16 var_8662_promoted_to_fp16 = const()[name = string("op_8662_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8668_cast_fp16 = pow(x = hidden_states_391_cast_fp16, y = var_8662_promoted_to_fp16)[name = string("op_8668_cast_fp16")];
            tensor<int32, [1]> variance_99_axes_0 = const()[name = string("variance_99_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_99_keep_dims_0 = const()[name = string("variance_99_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_99_cast_fp16 = reduce_mean(axes = variance_99_axes_0, keep_dims = variance_99_keep_dims_0, x = var_8668_cast_fp16)[name = string("variance_99_cast_fp16")];
            fp16 var_8671_to_fp16 = const()[name = string("op_8671_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8672_cast_fp16 = add(x = variance_99_cast_fp16, y = var_8671_to_fp16)[name = string("op_8672_cast_fp16")];
            fp32 var_8673_epsilon_0 = const()[name = string("op_8673_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8673_cast_fp16 = rsqrt(epsilon = var_8673_epsilon_0, x = var_8672_cast_fp16)[name = string("op_8673_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_395_cast_fp16 = mul(x = hidden_states_391_cast_fp16, y = var_8673_cast_fp16)[name = string("hidden_states_395_cast_fp16")];
            tensor<fp16, [1024]> layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(839000320)))];
            tensor<fp16, [1, 16, 1024]> input_199_cast_fp16 = mul(x = layers_24_post_attention_layernorm_weight_to_fp16, y = hidden_states_395_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_24_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_24_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(839002432)))];
            tensor<fp16, [1, 16, 4096]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_24_mlp_gate_proj_weight_to_fp16, x = input_199_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_8686_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_8686_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_24_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_24_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(847391104)))];
            tensor<fp16, [1, 16, 4096]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_24_mlp_up_proj_weight_to_fp16, x = input_199_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_203_cast_fp16 = mul(x = var_8686_cast_fp16, y = linear_173_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_24_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_24_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(855779776)))];
            tensor<fp16, [1, 16, 1024]> linear_174_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_24_mlp_down_proj_weight_to_fp16, x = input_203_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_399_cast_fp16 = add(x = hidden_states_391_cast_fp16, y = linear_174_cast_fp16)[name = string("hidden_states_399_cast_fp16")];
            tensor<int32, [4]> lkc_101_begin_0 = const()[name = string("lkc_101_begin_0"), val = tensor<int32, [4]>([0, 25600, 0, 0])];
            tensor<int32, [4]> lkc_101_end_0 = const()[name = string("lkc_101_end_0"), val = tensor<int32, [4]>([1, 26624, 1, 1024])];
            tensor<bool, [4]> lkc_101_end_mask_0 = const()[name = string("lkc_101_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_101_cast_fp16 = slice_by_index(begin = lkc_101_begin_0, end = lkc_101_end_0, end_mask = lkc_101_end_mask_0, x = key_cache)[name = string("lkc_101_cast_fp16")];
            tensor<int32, [4]> lvc_101_begin_0 = const()[name = string("lvc_101_begin_0"), val = tensor<int32, [4]>([0, 25600, 0, 0])];
            tensor<int32, [4]> lvc_101_end_0 = const()[name = string("lvc_101_end_0"), val = tensor<int32, [4]>([1, 26624, 1, 1024])];
            tensor<bool, [4]> lvc_101_end_mask_0 = const()[name = string("lvc_101_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_101_cast_fp16 = slice_by_index(begin = lvc_101_begin_0, end = lvc_101_end_0, end_mask = lvc_101_end_mask_0, x = value_cache)[name = string("lvc_101_cast_fp16")];
            fp16 var_8717_promoted_to_fp16 = const()[name = string("op_8717_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8723_cast_fp16 = pow(x = hidden_states_399_cast_fp16, y = var_8717_promoted_to_fp16)[name = string("op_8723_cast_fp16")];
            tensor<int32, [1]> variance_101_axes_0 = const()[name = string("variance_101_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_101_keep_dims_0 = const()[name = string("variance_101_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_101_cast_fp16 = reduce_mean(axes = variance_101_axes_0, keep_dims = variance_101_keep_dims_0, x = var_8723_cast_fp16)[name = string("variance_101_cast_fp16")];
            fp16 var_8726_to_fp16 = const()[name = string("op_8726_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8727_cast_fp16 = add(x = variance_101_cast_fp16, y = var_8726_to_fp16)[name = string("op_8727_cast_fp16")];
            fp32 var_8728_epsilon_0 = const()[name = string("op_8728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8728_cast_fp16 = rsqrt(epsilon = var_8728_epsilon_0, x = var_8727_cast_fp16)[name = string("op_8728_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_403_cast_fp16 = mul(x = hidden_states_399_cast_fp16, y = var_8728_cast_fp16)[name = string("hidden_states_403_cast_fp16")];
            tensor<fp16, [1024]> layers_25_input_layernorm_weight_to_fp16 = const()[name = string("layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(864168448)))];
            tensor<fp16, [1, 16, 1024]> hidden_101_cast_fp16 = mul(x = layers_25_input_layernorm_weight_to_fp16, y = hidden_states_403_cast_fp16)[name = string("hidden_101_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(864170560)))];
            tensor<fp16, [1, 16, 1024]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_q_proj_weight_to_fp16, x = hidden_101_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> var_8752 = const()[name = string("op_8752"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_151_cast_fp16 = reshape(shape = var_8752, x = linear_175_cast_fp16)[name = string("q_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(866267776)))];
            tensor<fp16, [1, 16, 1024]> linear_176_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_k_proj_weight_to_fp16, x = hidden_101_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> var_8759 = const()[name = string("op_8759"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_101_cast_fp16 = reshape(shape = var_8759, x = linear_176_cast_fp16)[name = string("k_101_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(868364992)))];
            tensor<fp16, [1, 16, 1024]> linear_177_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_v_proj_weight_to_fp16, x = hidden_101_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> var_8766 = const()[name = string("op_8766"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_51_cast_fp16 = reshape(shape = var_8766, x = linear_177_cast_fp16)[name = string("v_51_cast_fp16")];
            tensor<int32, [4]> q_153_perm_0 = const()[name = string("q_153_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_103_perm_0 = const()[name = string("k_103_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_25_perm_0 = const()[name = string("transpose_25_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_153_cast_fp16 = transpose(perm = q_153_perm_0, x = q_151_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 64]> var_8779_cast_fp16 = mul(x = q_153_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8779_cast_fp16")];
            tensor<int32, [4]> x1_101_begin_0 = const()[name = string("x1_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_101_end_0 = const()[name = string("x1_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_101_end_mask_0 = const()[name = string("x1_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_101_cast_fp16 = slice_by_index(begin = x1_101_begin_0, end = x1_101_end_0, end_mask = x1_101_end_mask_0, x = q_153_cast_fp16)[name = string("x1_101_cast_fp16")];
            tensor<int32, [4]> x2_101_begin_0 = const()[name = string("x2_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_101_end_0 = const()[name = string("x2_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_101_end_mask_0 = const()[name = string("x2_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_101_cast_fp16 = slice_by_index(begin = x2_101_begin_0, end = x2_101_end_0, end_mask = x2_101_end_mask_0, x = q_153_cast_fp16)[name = string("x2_101_cast_fp16")];
            fp16 const_307_promoted_to_fp16 = const()[name = string("const_307_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8800_cast_fp16 = mul(x = x2_101_cast_fp16, y = const_307_promoted_to_fp16)[name = string("op_8800_cast_fp16")];
            int32 var_8802 = const()[name = string("op_8802"), val = int32(-1)];
            bool var_8803_interleave_0 = const()[name = string("op_8803_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8803_cast_fp16 = concat(axis = var_8802, interleave = var_8803_interleave_0, values = (var_8800_cast_fp16, x1_101_cast_fp16))[name = string("op_8803_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8806_cast_fp16 = mul(x = var_8803_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8806_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_155_cast_fp16 = add(x = var_8779_cast_fp16, y = var_8806_cast_fp16)[name = string("q_155_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_103_cast_fp16 = transpose(perm = k_103_perm_0, x = k_101_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 16, 16, 64]> var_8811_cast_fp16 = mul(x = k_103_cast_fp16, y = var_1078_cast_fp16)[name = string("op_8811_cast_fp16")];
            tensor<int32, [4]> x1_103_begin_0 = const()[name = string("x1_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_103_end_0 = const()[name = string("x1_103_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_103_end_mask_0 = const()[name = string("x1_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_103_cast_fp16 = slice_by_index(begin = x1_103_begin_0, end = x1_103_end_0, end_mask = x1_103_end_mask_0, x = k_103_cast_fp16)[name = string("x1_103_cast_fp16")];
            tensor<int32, [4]> x2_103_begin_0 = const()[name = string("x2_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_103_end_0 = const()[name = string("x2_103_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_103_end_mask_0 = const()[name = string("x2_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_103_cast_fp16 = slice_by_index(begin = x2_103_begin_0, end = x2_103_end_0, end_mask = x2_103_end_mask_0, x = k_103_cast_fp16)[name = string("x2_103_cast_fp16")];
            fp16 const_310_promoted_to_fp16 = const()[name = string("const_310_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_8832_cast_fp16 = mul(x = x2_103_cast_fp16, y = const_310_promoted_to_fp16)[name = string("op_8832_cast_fp16")];
            int32 var_8834 = const()[name = string("op_8834"), val = int32(-1)];
            bool var_8835_interleave_0 = const()[name = string("op_8835_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_8835_cast_fp16 = concat(axis = var_8834, interleave = var_8835_interleave_0, values = (var_8832_cast_fp16, x1_103_cast_fp16))[name = string("op_8835_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8838_cast_fp16 = mul(x = var_8835_cast_fp16, y = var_1105_cast_fp16)[name = string("op_8838_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_101_cast_fp16 = add(x = var_8811_cast_fp16, y = var_8838_cast_fp16)[name = string("x_101_cast_fp16")];
            tensor<int32, [4]> var_8845 = const()[name = string("op_8845"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_8857 = const()[name = string("op_8857"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_8846_cast_fp16 = transpose(perm = var_8845, x = x_101_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 1024, 1, 16]> var_8858_cast_fp16 = reshape(shape = var_8857, x = var_8846_cast_fp16)[name = string("op_8858_cast_fp16")];
            tensor<int32, [1]> var_8860_axes_0 = const()[name = string("op_8860_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8860_cast_fp16 = squeeze(axes = var_8860_axes_0, x = var_8858_cast_fp16)[name = string("op_8860_cast_fp16")];
            bool var_8861_transpose_x_0 = const()[name = string("op_8861_transpose_x_0"), val = bool(false)];
            bool var_8861_transpose_y_0 = const()[name = string("op_8861_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8861_cast_fp16 = matmul(transpose_x = var_8861_transpose_x_0, transpose_y = var_8861_transpose_y_0, x = var_8860_cast_fp16, y = block_cache_map)[name = string("op_8861_cast_fp16")];
            tensor<int32, [1]> nk_full_51_axes_0 = const()[name = string("nk_full_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_51_cast_fp16 = expand_dims(axes = nk_full_51_axes_0, x = var_8861_cast_fp16)[name = string("nk_full_51_cast_fp16")];
            tensor<int32, [4]> var_8880 = const()[name = string("op_8880"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_25_cast_fp16 = transpose(perm = transpose_25_perm_0, x = v_51_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 1024, 1, 16]> var_8881_cast_fp16 = reshape(shape = var_8880, x = transpose_25_cast_fp16)[name = string("op_8881_cast_fp16")];
            tensor<int32, [1]> var_8883_axes_0 = const()[name = string("op_8883_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_8883_cast_fp16 = squeeze(axes = var_8883_axes_0, x = var_8881_cast_fp16)[name = string("op_8883_cast_fp16")];
            bool var_8884_transpose_x_0 = const()[name = string("op_8884_transpose_x_0"), val = bool(false)];
            bool var_8884_transpose_y_0 = const()[name = string("op_8884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_8884_cast_fp16 = matmul(transpose_x = var_8884_transpose_x_0, transpose_y = var_8884_transpose_y_0, x = var_8883_cast_fp16, y = block_cache_map)[name = string("op_8884_cast_fp16")];
            tensor<int32, [1]> nv_full_51_axes_0 = const()[name = string("nv_full_51_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_51_cast_fp16 = expand_dims(axes = nv_full_51_axes_0, x = var_8884_cast_fp16)[name = string("nv_full_51_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8890_cast_fp16 = mul(x = lkc_101_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8890_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8891_cast_fp16 = mul(x = nk_full_51_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8891_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_103_cast_fp16 = add(x = var_8890_cast_fp16, y = var_8891_cast_fp16)[name = string("lkc_103_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8897_cast_fp16 = mul(x = lvc_101_cast_fp16, y = var_1189_cast_fp16)[name = string("op_8897_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_8898_cast_fp16 = mul(x = nv_full_51_cast_fp16, y = update_mask_cast_fp16)[name = string("op_8898_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_103_cast_fp16 = add(x = var_8897_cast_fp16, y = var_8898_cast_fp16)[name = string("lvc_103_cast_fp16")];
            tensor<int32, [1]> var_8902_axes_0 = const()[name = string("op_8902_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8902_cast_fp16 = squeeze(axes = var_8902_axes_0, x = lkc_103_cast_fp16)[name = string("op_8902_cast_fp16")];
            tensor<int32, [4]> var_8906 = const()[name = string("op_8906"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_51_cast_fp16 = reshape(shape = var_8906, x = var_8902_cast_fp16)[name = string("kc_51_cast_fp16")];
            tensor<int32, [1]> var_8909_axes_0 = const()[name = string("op_8909_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_8909_cast_fp16 = squeeze(axes = var_8909_axes_0, x = lvc_103_cast_fp16)[name = string("op_8909_cast_fp16")];
            tensor<int32, [4]> var_8913 = const()[name = string("op_8913"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_51_cast_fp16 = reshape(shape = var_8913, x = var_8909_cast_fp16)[name = string("vc_51_cast_fp16")];
            bool var_8915_transpose_x_0 = const()[name = string("op_8915_transpose_x_0"), val = bool(false)];
            bool var_8915_transpose_y_0 = const()[name = string("op_8915_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_8915_cast_fp16 = matmul(transpose_x = var_8915_transpose_x_0, transpose_y = var_8915_transpose_y_0, x = q_155_cast_fp16, y = kc_51_cast_fp16)[name = string("op_8915_cast_fp16")];
            fp16 _inversed_scores_151_y_0_to_fp16 = const()[name = string("_inversed_scores_151_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_151_cast_fp16 = mul(x = var_8915_cast_fp16, y = _inversed_scores_151_y_0_to_fp16)[name = string("_inversed_scores_151_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_153_cast_fp16 = add(x = _inversed_scores_151_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_153_cast_fp16")];
            int32 var_8944 = const()[name = string("op_8944"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_8946_cast_fp16 = softmax(axis = var_8944, x = scores_153_cast_fp16)[name = string("op_8946_cast_fp16")];
            bool attn_out_101_transpose_x_1 = const()[name = string("attn_out_101_transpose_x_1"), val = bool(false)];
            bool attn_out_101_transpose_y_1 = const()[name = string("attn_out_101_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_101_cast_fp16 = matmul(transpose_x = attn_out_101_transpose_x_1, transpose_y = attn_out_101_transpose_y_1, x = var_8946_cast_fp16, y = vc_51_cast_fp16)[name = string("attn_out_101_cast_fp16")];
            tensor<int32, [4]> var_8958_perm_0 = const()[name = string("op_8958_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8960 = const()[name = string("op_8960"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_8958_cast_fp16 = transpose(perm = var_8958_perm_0, x = attn_out_101_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 1024]> input_205_cast_fp16 = reshape(shape = var_8960, x = var_8958_cast_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_25_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_25_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(870462208)))];
            tensor<fp16, [1, 16, 1024]> linear_178_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_self_attn_o_proj_weight_to_fp16, x = input_205_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_407_cast_fp16 = add(x = hidden_states_399_cast_fp16, y = linear_178_cast_fp16)[name = string("hidden_states_407_cast_fp16")];
            fp16 var_8970_promoted_to_fp16 = const()[name = string("op_8970_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8976_cast_fp16 = pow(x = hidden_states_407_cast_fp16, y = var_8970_promoted_to_fp16)[name = string("op_8976_cast_fp16")];
            tensor<int32, [1]> variance_103_axes_0 = const()[name = string("variance_103_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_103_keep_dims_0 = const()[name = string("variance_103_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_103_cast_fp16 = reduce_mean(axes = variance_103_axes_0, keep_dims = variance_103_keep_dims_0, x = var_8976_cast_fp16)[name = string("variance_103_cast_fp16")];
            fp16 var_8979_to_fp16 = const()[name = string("op_8979_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_8980_cast_fp16 = add(x = variance_103_cast_fp16, y = var_8979_to_fp16)[name = string("op_8980_cast_fp16")];
            fp32 var_8981_epsilon_0 = const()[name = string("op_8981_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8981_cast_fp16 = rsqrt(epsilon = var_8981_epsilon_0, x = var_8980_cast_fp16)[name = string("op_8981_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_411_cast_fp16 = mul(x = hidden_states_407_cast_fp16, y = var_8981_cast_fp16)[name = string("hidden_states_411_cast_fp16")];
            tensor<fp16, [1024]> layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872559424)))];
            tensor<fp16, [1, 16, 1024]> input_207_cast_fp16 = mul(x = layers_25_post_attention_layernorm_weight_to_fp16, y = hidden_states_411_cast_fp16)[name = string("input_207_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_25_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_25_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(872561536)))];
            tensor<fp16, [1, 16, 4096]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_25_mlp_gate_proj_weight_to_fp16, x = input_207_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_8994_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_8994_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_25_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_25_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(880950208)))];
            tensor<fp16, [1, 16, 4096]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_25_mlp_up_proj_weight_to_fp16, x = input_207_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_211_cast_fp16 = mul(x = var_8994_cast_fp16, y = linear_180_cast_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_25_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_25_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(889338880)))];
            tensor<fp16, [1, 16, 1024]> linear_181_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_25_mlp_down_proj_weight_to_fp16, x = input_211_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_415_cast_fp16 = add(x = hidden_states_407_cast_fp16, y = linear_181_cast_fp16)[name = string("hidden_states_415_cast_fp16")];
            tensor<int32, [4]> lkc_105_begin_0 = const()[name = string("lkc_105_begin_0"), val = tensor<int32, [4]>([0, 26624, 0, 0])];
            tensor<int32, [4]> lkc_105_end_0 = const()[name = string("lkc_105_end_0"), val = tensor<int32, [4]>([1, 27648, 1, 1024])];
            tensor<bool, [4]> lkc_105_end_mask_0 = const()[name = string("lkc_105_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_105_cast_fp16 = slice_by_index(begin = lkc_105_begin_0, end = lkc_105_end_0, end_mask = lkc_105_end_mask_0, x = key_cache)[name = string("lkc_105_cast_fp16")];
            tensor<int32, [4]> lvc_105_begin_0 = const()[name = string("lvc_105_begin_0"), val = tensor<int32, [4]>([0, 26624, 0, 0])];
            tensor<int32, [4]> lvc_105_end_0 = const()[name = string("lvc_105_end_0"), val = tensor<int32, [4]>([1, 27648, 1, 1024])];
            tensor<bool, [4]> lvc_105_end_mask_0 = const()[name = string("lvc_105_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_105_cast_fp16 = slice_by_index(begin = lvc_105_begin_0, end = lvc_105_end_0, end_mask = lvc_105_end_mask_0, x = value_cache)[name = string("lvc_105_cast_fp16")];
            fp16 var_9025_promoted_to_fp16 = const()[name = string("op_9025_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9031_cast_fp16 = pow(x = hidden_states_415_cast_fp16, y = var_9025_promoted_to_fp16)[name = string("op_9031_cast_fp16")];
            tensor<int32, [1]> variance_105_axes_0 = const()[name = string("variance_105_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_105_keep_dims_0 = const()[name = string("variance_105_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_105_cast_fp16 = reduce_mean(axes = variance_105_axes_0, keep_dims = variance_105_keep_dims_0, x = var_9031_cast_fp16)[name = string("variance_105_cast_fp16")];
            fp16 var_9034_to_fp16 = const()[name = string("op_9034_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9035_cast_fp16 = add(x = variance_105_cast_fp16, y = var_9034_to_fp16)[name = string("op_9035_cast_fp16")];
            fp32 var_9036_epsilon_0 = const()[name = string("op_9036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9036_cast_fp16 = rsqrt(epsilon = var_9036_epsilon_0, x = var_9035_cast_fp16)[name = string("op_9036_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_419_cast_fp16 = mul(x = hidden_states_415_cast_fp16, y = var_9036_cast_fp16)[name = string("hidden_states_419_cast_fp16")];
            tensor<fp16, [1024]> layers_26_input_layernorm_weight_to_fp16 = const()[name = string("layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897727552)))];
            tensor<fp16, [1, 16, 1024]> hidden_105_cast_fp16 = mul(x = layers_26_input_layernorm_weight_to_fp16, y = hidden_states_419_cast_fp16)[name = string("hidden_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(897729664)))];
            tensor<fp16, [1, 16, 1024]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_q_proj_weight_to_fp16, x = hidden_105_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_9060 = const()[name = string("op_9060"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_157_cast_fp16 = reshape(shape = var_9060, x = linear_182_cast_fp16)[name = string("q_157_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(899826880)))];
            tensor<fp16, [1, 16, 1024]> linear_183_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_k_proj_weight_to_fp16, x = hidden_105_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_9067 = const()[name = string("op_9067"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_105_cast_fp16 = reshape(shape = var_9067, x = linear_183_cast_fp16)[name = string("k_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(901924096)))];
            tensor<fp16, [1, 16, 1024]> linear_184_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_v_proj_weight_to_fp16, x = hidden_105_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_9074 = const()[name = string("op_9074"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_53_cast_fp16 = reshape(shape = var_9074, x = linear_184_cast_fp16)[name = string("v_53_cast_fp16")];
            tensor<int32, [4]> q_159_perm_0 = const()[name = string("q_159_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_107_perm_0 = const()[name = string("k_107_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_26_perm_0 = const()[name = string("transpose_26_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_159_cast_fp16 = transpose(perm = q_159_perm_0, x = q_157_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 64]> var_9087_cast_fp16 = mul(x = q_159_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9087_cast_fp16")];
            tensor<int32, [4]> x1_105_begin_0 = const()[name = string("x1_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_105_end_0 = const()[name = string("x1_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_105_end_mask_0 = const()[name = string("x1_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_105_cast_fp16 = slice_by_index(begin = x1_105_begin_0, end = x1_105_end_0, end_mask = x1_105_end_mask_0, x = q_159_cast_fp16)[name = string("x1_105_cast_fp16")];
            tensor<int32, [4]> x2_105_begin_0 = const()[name = string("x2_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_105_end_0 = const()[name = string("x2_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_105_end_mask_0 = const()[name = string("x2_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_105_cast_fp16 = slice_by_index(begin = x2_105_begin_0, end = x2_105_end_0, end_mask = x2_105_end_mask_0, x = q_159_cast_fp16)[name = string("x2_105_cast_fp16")];
            fp16 const_319_promoted_to_fp16 = const()[name = string("const_319_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9108_cast_fp16 = mul(x = x2_105_cast_fp16, y = const_319_promoted_to_fp16)[name = string("op_9108_cast_fp16")];
            int32 var_9110 = const()[name = string("op_9110"), val = int32(-1)];
            bool var_9111_interleave_0 = const()[name = string("op_9111_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9111_cast_fp16 = concat(axis = var_9110, interleave = var_9111_interleave_0, values = (var_9108_cast_fp16, x1_105_cast_fp16))[name = string("op_9111_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9114_cast_fp16 = mul(x = var_9111_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9114_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_161_cast_fp16 = add(x = var_9087_cast_fp16, y = var_9114_cast_fp16)[name = string("q_161_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_107_cast_fp16 = transpose(perm = k_107_perm_0, x = k_105_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 16, 16, 64]> var_9119_cast_fp16 = mul(x = k_107_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9119_cast_fp16")];
            tensor<int32, [4]> x1_107_begin_0 = const()[name = string("x1_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_107_end_0 = const()[name = string("x1_107_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_107_end_mask_0 = const()[name = string("x1_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_107_cast_fp16 = slice_by_index(begin = x1_107_begin_0, end = x1_107_end_0, end_mask = x1_107_end_mask_0, x = k_107_cast_fp16)[name = string("x1_107_cast_fp16")];
            tensor<int32, [4]> x2_107_begin_0 = const()[name = string("x2_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_107_end_0 = const()[name = string("x2_107_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_107_end_mask_0 = const()[name = string("x2_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_107_cast_fp16 = slice_by_index(begin = x2_107_begin_0, end = x2_107_end_0, end_mask = x2_107_end_mask_0, x = k_107_cast_fp16)[name = string("x2_107_cast_fp16")];
            fp16 const_322_promoted_to_fp16 = const()[name = string("const_322_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9140_cast_fp16 = mul(x = x2_107_cast_fp16, y = const_322_promoted_to_fp16)[name = string("op_9140_cast_fp16")];
            int32 var_9142 = const()[name = string("op_9142"), val = int32(-1)];
            bool var_9143_interleave_0 = const()[name = string("op_9143_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9143_cast_fp16 = concat(axis = var_9142, interleave = var_9143_interleave_0, values = (var_9140_cast_fp16, x1_107_cast_fp16))[name = string("op_9143_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9146_cast_fp16 = mul(x = var_9143_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9146_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_105_cast_fp16 = add(x = var_9119_cast_fp16, y = var_9146_cast_fp16)[name = string("x_105_cast_fp16")];
            tensor<int32, [4]> var_9153 = const()[name = string("op_9153"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9165 = const()[name = string("op_9165"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_9154_cast_fp16 = transpose(perm = var_9153, x = x_105_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 1024, 1, 16]> var_9166_cast_fp16 = reshape(shape = var_9165, x = var_9154_cast_fp16)[name = string("op_9166_cast_fp16")];
            tensor<int32, [1]> var_9168_axes_0 = const()[name = string("op_9168_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9168_cast_fp16 = squeeze(axes = var_9168_axes_0, x = var_9166_cast_fp16)[name = string("op_9168_cast_fp16")];
            bool var_9169_transpose_x_0 = const()[name = string("op_9169_transpose_x_0"), val = bool(false)];
            bool var_9169_transpose_y_0 = const()[name = string("op_9169_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9169_cast_fp16 = matmul(transpose_x = var_9169_transpose_x_0, transpose_y = var_9169_transpose_y_0, x = var_9168_cast_fp16, y = block_cache_map)[name = string("op_9169_cast_fp16")];
            tensor<int32, [1]> nk_full_53_axes_0 = const()[name = string("nk_full_53_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_53_cast_fp16 = expand_dims(axes = nk_full_53_axes_0, x = var_9169_cast_fp16)[name = string("nk_full_53_cast_fp16")];
            tensor<int32, [4]> var_9188 = const()[name = string("op_9188"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_26_cast_fp16 = transpose(perm = transpose_26_perm_0, x = v_53_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 1024, 1, 16]> var_9189_cast_fp16 = reshape(shape = var_9188, x = transpose_26_cast_fp16)[name = string("op_9189_cast_fp16")];
            tensor<int32, [1]> var_9191_axes_0 = const()[name = string("op_9191_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9191_cast_fp16 = squeeze(axes = var_9191_axes_0, x = var_9189_cast_fp16)[name = string("op_9191_cast_fp16")];
            bool var_9192_transpose_x_0 = const()[name = string("op_9192_transpose_x_0"), val = bool(false)];
            bool var_9192_transpose_y_0 = const()[name = string("op_9192_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9192_cast_fp16 = matmul(transpose_x = var_9192_transpose_x_0, transpose_y = var_9192_transpose_y_0, x = var_9191_cast_fp16, y = block_cache_map)[name = string("op_9192_cast_fp16")];
            tensor<int32, [1]> nv_full_53_axes_0 = const()[name = string("nv_full_53_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_53_cast_fp16 = expand_dims(axes = nv_full_53_axes_0, x = var_9192_cast_fp16)[name = string("nv_full_53_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9198_cast_fp16 = mul(x = lkc_105_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9198_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9199_cast_fp16 = mul(x = nk_full_53_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9199_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_107_cast_fp16 = add(x = var_9198_cast_fp16, y = var_9199_cast_fp16)[name = string("lkc_107_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9205_cast_fp16 = mul(x = lvc_105_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9205_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9206_cast_fp16 = mul(x = nv_full_53_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9206_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_107_cast_fp16 = add(x = var_9205_cast_fp16, y = var_9206_cast_fp16)[name = string("lvc_107_cast_fp16")];
            tensor<int32, [1]> var_9210_axes_0 = const()[name = string("op_9210_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9210_cast_fp16 = squeeze(axes = var_9210_axes_0, x = lkc_107_cast_fp16)[name = string("op_9210_cast_fp16")];
            tensor<int32, [4]> var_9214 = const()[name = string("op_9214"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_53_cast_fp16 = reshape(shape = var_9214, x = var_9210_cast_fp16)[name = string("kc_53_cast_fp16")];
            tensor<int32, [1]> var_9217_axes_0 = const()[name = string("op_9217_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9217_cast_fp16 = squeeze(axes = var_9217_axes_0, x = lvc_107_cast_fp16)[name = string("op_9217_cast_fp16")];
            tensor<int32, [4]> var_9221 = const()[name = string("op_9221"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_53_cast_fp16 = reshape(shape = var_9221, x = var_9217_cast_fp16)[name = string("vc_53_cast_fp16")];
            bool var_9223_transpose_x_0 = const()[name = string("op_9223_transpose_x_0"), val = bool(false)];
            bool var_9223_transpose_y_0 = const()[name = string("op_9223_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_9223_cast_fp16 = matmul(transpose_x = var_9223_transpose_x_0, transpose_y = var_9223_transpose_y_0, x = q_161_cast_fp16, y = kc_53_cast_fp16)[name = string("op_9223_cast_fp16")];
            fp16 _inversed_scores_157_y_0_to_fp16 = const()[name = string("_inversed_scores_157_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_157_cast_fp16 = mul(x = var_9223_cast_fp16, y = _inversed_scores_157_y_0_to_fp16)[name = string("_inversed_scores_157_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_159_cast_fp16 = add(x = _inversed_scores_157_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_159_cast_fp16")];
            int32 var_9252 = const()[name = string("op_9252"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_9254_cast_fp16 = softmax(axis = var_9252, x = scores_159_cast_fp16)[name = string("op_9254_cast_fp16")];
            bool attn_out_105_transpose_x_1 = const()[name = string("attn_out_105_transpose_x_1"), val = bool(false)];
            bool attn_out_105_transpose_y_1 = const()[name = string("attn_out_105_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_105_cast_fp16 = matmul(transpose_x = attn_out_105_transpose_x_1, transpose_y = attn_out_105_transpose_y_1, x = var_9254_cast_fp16, y = vc_53_cast_fp16)[name = string("attn_out_105_cast_fp16")];
            tensor<int32, [4]> var_9266_perm_0 = const()[name = string("op_9266_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9268 = const()[name = string("op_9268"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_9266_cast_fp16 = transpose(perm = var_9266_perm_0, x = attn_out_105_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 1024]> input_213_cast_fp16 = reshape(shape = var_9268, x = var_9266_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_26_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_26_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(904021312)))];
            tensor<fp16, [1, 16, 1024]> linear_185_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_self_attn_o_proj_weight_to_fp16, x = input_213_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_423_cast_fp16 = add(x = hidden_states_415_cast_fp16, y = linear_185_cast_fp16)[name = string("hidden_states_423_cast_fp16")];
            fp16 var_9278_promoted_to_fp16 = const()[name = string("op_9278_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9284_cast_fp16 = pow(x = hidden_states_423_cast_fp16, y = var_9278_promoted_to_fp16)[name = string("op_9284_cast_fp16")];
            tensor<int32, [1]> variance_107_axes_0 = const()[name = string("variance_107_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_107_keep_dims_0 = const()[name = string("variance_107_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_107_cast_fp16 = reduce_mean(axes = variance_107_axes_0, keep_dims = variance_107_keep_dims_0, x = var_9284_cast_fp16)[name = string("variance_107_cast_fp16")];
            fp16 var_9287_to_fp16 = const()[name = string("op_9287_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9288_cast_fp16 = add(x = variance_107_cast_fp16, y = var_9287_to_fp16)[name = string("op_9288_cast_fp16")];
            fp32 var_9289_epsilon_0 = const()[name = string("op_9289_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9289_cast_fp16 = rsqrt(epsilon = var_9289_epsilon_0, x = var_9288_cast_fp16)[name = string("op_9289_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_427_cast_fp16 = mul(x = hidden_states_423_cast_fp16, y = var_9289_cast_fp16)[name = string("hidden_states_427_cast_fp16")];
            tensor<fp16, [1024]> layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906118528)))];
            tensor<fp16, [1, 16, 1024]> input_215_cast_fp16 = mul(x = layers_26_post_attention_layernorm_weight_to_fp16, y = hidden_states_427_cast_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_26_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_26_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(906120640)))];
            tensor<fp16, [1, 16, 4096]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_26_mlp_gate_proj_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_9302_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_9302_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_26_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_26_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(914509312)))];
            tensor<fp16, [1, 16, 4096]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_26_mlp_up_proj_weight_to_fp16, x = input_215_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_219_cast_fp16 = mul(x = var_9302_cast_fp16, y = linear_187_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_26_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_26_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(922897984)))];
            tensor<fp16, [1, 16, 1024]> linear_188_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_26_mlp_down_proj_weight_to_fp16, x = input_219_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_431_cast_fp16 = add(x = hidden_states_423_cast_fp16, y = linear_188_cast_fp16)[name = string("hidden_states_431_cast_fp16")];
            tensor<int32, [4]> lkc_109_begin_0 = const()[name = string("lkc_109_begin_0"), val = tensor<int32, [4]>([0, 27648, 0, 0])];
            tensor<int32, [4]> lkc_109_end_0 = const()[name = string("lkc_109_end_0"), val = tensor<int32, [4]>([1, 28672, 1, 1024])];
            tensor<bool, [4]> lkc_109_end_mask_0 = const()[name = string("lkc_109_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_109_cast_fp16 = slice_by_index(begin = lkc_109_begin_0, end = lkc_109_end_0, end_mask = lkc_109_end_mask_0, x = key_cache)[name = string("lkc_109_cast_fp16")];
            tensor<int32, [4]> lvc_109_begin_0 = const()[name = string("lvc_109_begin_0"), val = tensor<int32, [4]>([0, 27648, 0, 0])];
            tensor<int32, [4]> lvc_109_end_0 = const()[name = string("lvc_109_end_0"), val = tensor<int32, [4]>([1, 28672, 1, 1024])];
            tensor<bool, [4]> lvc_109_end_mask_0 = const()[name = string("lvc_109_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_109_cast_fp16 = slice_by_index(begin = lvc_109_begin_0, end = lvc_109_end_0, end_mask = lvc_109_end_mask_0, x = value_cache)[name = string("lvc_109_cast_fp16")];
            fp16 var_9333_promoted_to_fp16 = const()[name = string("op_9333_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9339_cast_fp16 = pow(x = hidden_states_431_cast_fp16, y = var_9333_promoted_to_fp16)[name = string("op_9339_cast_fp16")];
            tensor<int32, [1]> variance_109_axes_0 = const()[name = string("variance_109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_109_keep_dims_0 = const()[name = string("variance_109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_109_cast_fp16 = reduce_mean(axes = variance_109_axes_0, keep_dims = variance_109_keep_dims_0, x = var_9339_cast_fp16)[name = string("variance_109_cast_fp16")];
            fp16 var_9342_to_fp16 = const()[name = string("op_9342_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9343_cast_fp16 = add(x = variance_109_cast_fp16, y = var_9342_to_fp16)[name = string("op_9343_cast_fp16")];
            fp32 var_9344_epsilon_0 = const()[name = string("op_9344_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9344_cast_fp16 = rsqrt(epsilon = var_9344_epsilon_0, x = var_9343_cast_fp16)[name = string("op_9344_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_435_cast_fp16 = mul(x = hidden_states_431_cast_fp16, y = var_9344_cast_fp16)[name = string("hidden_states_435_cast_fp16")];
            tensor<fp16, [1024]> layers_27_input_layernorm_weight_to_fp16 = const()[name = string("layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(931286656)))];
            tensor<fp16, [1, 16, 1024]> hidden_109_cast_fp16 = mul(x = layers_27_input_layernorm_weight_to_fp16, y = hidden_states_435_cast_fp16)[name = string("hidden_109_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(931288768)))];
            tensor<fp16, [1, 16, 1024]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_q_proj_weight_to_fp16, x = hidden_109_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<int32, [4]> var_9368 = const()[name = string("op_9368"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_163_cast_fp16 = reshape(shape = var_9368, x = linear_189_cast_fp16)[name = string("q_163_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(933385984)))];
            tensor<fp16, [1, 16, 1024]> linear_190_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_k_proj_weight_to_fp16, x = hidden_109_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9375 = const()[name = string("op_9375"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_109_cast_fp16 = reshape(shape = var_9375, x = linear_190_cast_fp16)[name = string("k_109_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(935483200)))];
            tensor<fp16, [1, 16, 1024]> linear_191_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_v_proj_weight_to_fp16, x = hidden_109_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9382 = const()[name = string("op_9382"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_55_cast_fp16 = reshape(shape = var_9382, x = linear_191_cast_fp16)[name = string("v_55_cast_fp16")];
            tensor<int32, [4]> q_165_perm_0 = const()[name = string("q_165_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_111_perm_0 = const()[name = string("k_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_27_perm_0 = const()[name = string("transpose_27_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_165_cast_fp16 = transpose(perm = q_165_perm_0, x = q_163_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 64]> var_9395_cast_fp16 = mul(x = q_165_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9395_cast_fp16")];
            tensor<int32, [4]> x1_109_begin_0 = const()[name = string("x1_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_109_end_0 = const()[name = string("x1_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_109_end_mask_0 = const()[name = string("x1_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_109_cast_fp16 = slice_by_index(begin = x1_109_begin_0, end = x1_109_end_0, end_mask = x1_109_end_mask_0, x = q_165_cast_fp16)[name = string("x1_109_cast_fp16")];
            tensor<int32, [4]> x2_109_begin_0 = const()[name = string("x2_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_109_end_0 = const()[name = string("x2_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_109_end_mask_0 = const()[name = string("x2_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_109_cast_fp16 = slice_by_index(begin = x2_109_begin_0, end = x2_109_end_0, end_mask = x2_109_end_mask_0, x = q_165_cast_fp16)[name = string("x2_109_cast_fp16")];
            fp16 const_331_promoted_to_fp16 = const()[name = string("const_331_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9416_cast_fp16 = mul(x = x2_109_cast_fp16, y = const_331_promoted_to_fp16)[name = string("op_9416_cast_fp16")];
            int32 var_9418 = const()[name = string("op_9418"), val = int32(-1)];
            bool var_9419_interleave_0 = const()[name = string("op_9419_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9419_cast_fp16 = concat(axis = var_9418, interleave = var_9419_interleave_0, values = (var_9416_cast_fp16, x1_109_cast_fp16))[name = string("op_9419_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9422_cast_fp16 = mul(x = var_9419_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9422_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_167_cast_fp16 = add(x = var_9395_cast_fp16, y = var_9422_cast_fp16)[name = string("q_167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_111_cast_fp16 = transpose(perm = k_111_perm_0, x = k_109_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 16, 16, 64]> var_9427_cast_fp16 = mul(x = k_111_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9427_cast_fp16")];
            tensor<int32, [4]> x1_111_begin_0 = const()[name = string("x1_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_111_end_0 = const()[name = string("x1_111_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_111_end_mask_0 = const()[name = string("x1_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_111_cast_fp16 = slice_by_index(begin = x1_111_begin_0, end = x1_111_end_0, end_mask = x1_111_end_mask_0, x = k_111_cast_fp16)[name = string("x1_111_cast_fp16")];
            tensor<int32, [4]> x2_111_begin_0 = const()[name = string("x2_111_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_111_end_0 = const()[name = string("x2_111_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_111_end_mask_0 = const()[name = string("x2_111_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_111_cast_fp16 = slice_by_index(begin = x2_111_begin_0, end = x2_111_end_0, end_mask = x2_111_end_mask_0, x = k_111_cast_fp16)[name = string("x2_111_cast_fp16")];
            fp16 const_334_promoted_to_fp16 = const()[name = string("const_334_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9448_cast_fp16 = mul(x = x2_111_cast_fp16, y = const_334_promoted_to_fp16)[name = string("op_9448_cast_fp16")];
            int32 var_9450 = const()[name = string("op_9450"), val = int32(-1)];
            bool var_9451_interleave_0 = const()[name = string("op_9451_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9451_cast_fp16 = concat(axis = var_9450, interleave = var_9451_interleave_0, values = (var_9448_cast_fp16, x1_111_cast_fp16))[name = string("op_9451_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9454_cast_fp16 = mul(x = var_9451_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9454_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_109_cast_fp16 = add(x = var_9427_cast_fp16, y = var_9454_cast_fp16)[name = string("x_109_cast_fp16")];
            tensor<int32, [4]> var_9461 = const()[name = string("op_9461"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9473 = const()[name = string("op_9473"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_9462_cast_fp16 = transpose(perm = var_9461, x = x_109_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 1024, 1, 16]> var_9474_cast_fp16 = reshape(shape = var_9473, x = var_9462_cast_fp16)[name = string("op_9474_cast_fp16")];
            tensor<int32, [1]> var_9476_axes_0 = const()[name = string("op_9476_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9476_cast_fp16 = squeeze(axes = var_9476_axes_0, x = var_9474_cast_fp16)[name = string("op_9476_cast_fp16")];
            bool var_9477_transpose_x_0 = const()[name = string("op_9477_transpose_x_0"), val = bool(false)];
            bool var_9477_transpose_y_0 = const()[name = string("op_9477_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9477_cast_fp16 = matmul(transpose_x = var_9477_transpose_x_0, transpose_y = var_9477_transpose_y_0, x = var_9476_cast_fp16, y = block_cache_map)[name = string("op_9477_cast_fp16")];
            tensor<int32, [1]> nk_full_55_axes_0 = const()[name = string("nk_full_55_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_55_cast_fp16 = expand_dims(axes = nk_full_55_axes_0, x = var_9477_cast_fp16)[name = string("nk_full_55_cast_fp16")];
            tensor<int32, [4]> var_9496 = const()[name = string("op_9496"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_27_cast_fp16 = transpose(perm = transpose_27_perm_0, x = v_55_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 1024, 1, 16]> var_9497_cast_fp16 = reshape(shape = var_9496, x = transpose_27_cast_fp16)[name = string("op_9497_cast_fp16")];
            tensor<int32, [1]> var_9499_axes_0 = const()[name = string("op_9499_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9499_cast_fp16 = squeeze(axes = var_9499_axes_0, x = var_9497_cast_fp16)[name = string("op_9499_cast_fp16")];
            bool var_9500_transpose_x_0 = const()[name = string("op_9500_transpose_x_0"), val = bool(false)];
            bool var_9500_transpose_y_0 = const()[name = string("op_9500_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9500_cast_fp16 = matmul(transpose_x = var_9500_transpose_x_0, transpose_y = var_9500_transpose_y_0, x = var_9499_cast_fp16, y = block_cache_map)[name = string("op_9500_cast_fp16")];
            tensor<int32, [1]> nv_full_55_axes_0 = const()[name = string("nv_full_55_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_55_cast_fp16 = expand_dims(axes = nv_full_55_axes_0, x = var_9500_cast_fp16)[name = string("nv_full_55_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9506_cast_fp16 = mul(x = lkc_109_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9506_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9507_cast_fp16 = mul(x = nk_full_55_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9507_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_111_cast_fp16 = add(x = var_9506_cast_fp16, y = var_9507_cast_fp16)[name = string("lkc_111_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9513_cast_fp16 = mul(x = lvc_109_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9513_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9514_cast_fp16 = mul(x = nv_full_55_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9514_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_111_cast_fp16 = add(x = var_9513_cast_fp16, y = var_9514_cast_fp16)[name = string("lvc_111_cast_fp16")];
            tensor<int32, [1]> var_9518_axes_0 = const()[name = string("op_9518_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9518_cast_fp16 = squeeze(axes = var_9518_axes_0, x = lkc_111_cast_fp16)[name = string("op_9518_cast_fp16")];
            tensor<int32, [4]> var_9522 = const()[name = string("op_9522"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_55_cast_fp16 = reshape(shape = var_9522, x = var_9518_cast_fp16)[name = string("kc_55_cast_fp16")];
            tensor<int32, [1]> var_9525_axes_0 = const()[name = string("op_9525_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9525_cast_fp16 = squeeze(axes = var_9525_axes_0, x = lvc_111_cast_fp16)[name = string("op_9525_cast_fp16")];
            tensor<int32, [4]> var_9529 = const()[name = string("op_9529"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_55_cast_fp16 = reshape(shape = var_9529, x = var_9525_cast_fp16)[name = string("vc_55_cast_fp16")];
            bool var_9531_transpose_x_0 = const()[name = string("op_9531_transpose_x_0"), val = bool(false)];
            bool var_9531_transpose_y_0 = const()[name = string("op_9531_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_9531_cast_fp16 = matmul(transpose_x = var_9531_transpose_x_0, transpose_y = var_9531_transpose_y_0, x = q_167_cast_fp16, y = kc_55_cast_fp16)[name = string("op_9531_cast_fp16")];
            fp16 _inversed_scores_163_y_0_to_fp16 = const()[name = string("_inversed_scores_163_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_163_cast_fp16 = mul(x = var_9531_cast_fp16, y = _inversed_scores_163_y_0_to_fp16)[name = string("_inversed_scores_163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_165_cast_fp16 = add(x = _inversed_scores_163_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_165_cast_fp16")];
            int32 var_9560 = const()[name = string("op_9560"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_9562_cast_fp16 = softmax(axis = var_9560, x = scores_165_cast_fp16)[name = string("op_9562_cast_fp16")];
            bool attn_out_109_transpose_x_1 = const()[name = string("attn_out_109_transpose_x_1"), val = bool(false)];
            bool attn_out_109_transpose_y_1 = const()[name = string("attn_out_109_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_109_cast_fp16 = matmul(transpose_x = attn_out_109_transpose_x_1, transpose_y = attn_out_109_transpose_y_1, x = var_9562_cast_fp16, y = vc_55_cast_fp16)[name = string("attn_out_109_cast_fp16")];
            tensor<int32, [4]> var_9574_perm_0 = const()[name = string("op_9574_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9576 = const()[name = string("op_9576"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_9574_cast_fp16 = transpose(perm = var_9574_perm_0, x = attn_out_109_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 1024]> input_221_cast_fp16 = reshape(shape = var_9576, x = var_9574_cast_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_27_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_27_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(937580416)))];
            tensor<fp16, [1, 16, 1024]> linear_192_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_self_attn_o_proj_weight_to_fp16, x = input_221_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_439_cast_fp16 = add(x = hidden_states_431_cast_fp16, y = linear_192_cast_fp16)[name = string("hidden_states_439_cast_fp16")];
            fp16 var_9586_promoted_to_fp16 = const()[name = string("op_9586_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9592_cast_fp16 = pow(x = hidden_states_439_cast_fp16, y = var_9586_promoted_to_fp16)[name = string("op_9592_cast_fp16")];
            tensor<int32, [1]> variance_111_axes_0 = const()[name = string("variance_111_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_111_keep_dims_0 = const()[name = string("variance_111_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_111_cast_fp16 = reduce_mean(axes = variance_111_axes_0, keep_dims = variance_111_keep_dims_0, x = var_9592_cast_fp16)[name = string("variance_111_cast_fp16")];
            fp16 var_9595_to_fp16 = const()[name = string("op_9595_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9596_cast_fp16 = add(x = variance_111_cast_fp16, y = var_9595_to_fp16)[name = string("op_9596_cast_fp16")];
            fp32 var_9597_epsilon_0 = const()[name = string("op_9597_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9597_cast_fp16 = rsqrt(epsilon = var_9597_epsilon_0, x = var_9596_cast_fp16)[name = string("op_9597_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_443_cast_fp16 = mul(x = hidden_states_439_cast_fp16, y = var_9597_cast_fp16)[name = string("hidden_states_443_cast_fp16")];
            tensor<fp16, [1024]> layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(939677632)))];
            tensor<fp16, [1, 16, 1024]> input_223_cast_fp16 = mul(x = layers_27_post_attention_layernorm_weight_to_fp16, y = hidden_states_443_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_27_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(939679744)))];
            tensor<fp16, [1, 16, 4096]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_27_mlp_gate_proj_weight_to_fp16, x = input_223_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_9610_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_9610_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_27_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(948068416)))];
            tensor<fp16, [1, 16, 4096]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_27_mlp_up_proj_weight_to_fp16, x = input_223_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_227_cast_fp16 = mul(x = var_9610_cast_fp16, y = linear_194_cast_fp16)[name = string("input_227_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_27_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_27_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(956457088)))];
            tensor<fp16, [1, 16, 1024]> linear_195_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_27_mlp_down_proj_weight_to_fp16, x = input_227_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_447_cast_fp16 = add(x = hidden_states_439_cast_fp16, y = linear_195_cast_fp16)[name = string("hidden_states_447_cast_fp16")];
            tensor<int32, [4]> lkc_113_begin_0 = const()[name = string("lkc_113_begin_0"), val = tensor<int32, [4]>([0, 28672, 0, 0])];
            tensor<int32, [4]> lkc_113_end_0 = const()[name = string("lkc_113_end_0"), val = tensor<int32, [4]>([1, 29696, 1, 1024])];
            tensor<bool, [4]> lkc_113_end_mask_0 = const()[name = string("lkc_113_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_113_cast_fp16 = slice_by_index(begin = lkc_113_begin_0, end = lkc_113_end_0, end_mask = lkc_113_end_mask_0, x = key_cache)[name = string("lkc_113_cast_fp16")];
            tensor<int32, [4]> lvc_113_begin_0 = const()[name = string("lvc_113_begin_0"), val = tensor<int32, [4]>([0, 28672, 0, 0])];
            tensor<int32, [4]> lvc_113_end_0 = const()[name = string("lvc_113_end_0"), val = tensor<int32, [4]>([1, 29696, 1, 1024])];
            tensor<bool, [4]> lvc_113_end_mask_0 = const()[name = string("lvc_113_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_113_cast_fp16 = slice_by_index(begin = lvc_113_begin_0, end = lvc_113_end_0, end_mask = lvc_113_end_mask_0, x = value_cache)[name = string("lvc_113_cast_fp16")];
            fp16 var_9641_promoted_to_fp16 = const()[name = string("op_9641_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9647_cast_fp16 = pow(x = hidden_states_447_cast_fp16, y = var_9641_promoted_to_fp16)[name = string("op_9647_cast_fp16")];
            tensor<int32, [1]> variance_113_axes_0 = const()[name = string("variance_113_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_113_keep_dims_0 = const()[name = string("variance_113_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_113_cast_fp16 = reduce_mean(axes = variance_113_axes_0, keep_dims = variance_113_keep_dims_0, x = var_9647_cast_fp16)[name = string("variance_113_cast_fp16")];
            fp16 var_9650_to_fp16 = const()[name = string("op_9650_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9651_cast_fp16 = add(x = variance_113_cast_fp16, y = var_9650_to_fp16)[name = string("op_9651_cast_fp16")];
            fp32 var_9652_epsilon_0 = const()[name = string("op_9652_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9652_cast_fp16 = rsqrt(epsilon = var_9652_epsilon_0, x = var_9651_cast_fp16)[name = string("op_9652_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_451_cast_fp16 = mul(x = hidden_states_447_cast_fp16, y = var_9652_cast_fp16)[name = string("hidden_states_451_cast_fp16")];
            tensor<fp16, [1024]> layers_28_input_layernorm_weight_to_fp16 = const()[name = string("layers_28_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(964845760)))];
            tensor<fp16, [1, 16, 1024]> hidden_113_cast_fp16 = mul(x = layers_28_input_layernorm_weight_to_fp16, y = hidden_states_451_cast_fp16)[name = string("hidden_113_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_28_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(964847872)))];
            tensor<fp16, [1, 16, 1024]> linear_196_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_28_self_attn_q_proj_weight_to_fp16, x = hidden_113_cast_fp16)[name = string("linear_196_cast_fp16")];
            tensor<int32, [4]> var_9676 = const()[name = string("op_9676"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_169_cast_fp16 = reshape(shape = var_9676, x = linear_196_cast_fp16)[name = string("q_169_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_28_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(966945088)))];
            tensor<fp16, [1, 16, 1024]> linear_197_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_28_self_attn_k_proj_weight_to_fp16, x = hidden_113_cast_fp16)[name = string("linear_197_cast_fp16")];
            tensor<int32, [4]> var_9683 = const()[name = string("op_9683"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_113_cast_fp16 = reshape(shape = var_9683, x = linear_197_cast_fp16)[name = string("k_113_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_28_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(969042304)))];
            tensor<fp16, [1, 16, 1024]> linear_198_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_28_self_attn_v_proj_weight_to_fp16, x = hidden_113_cast_fp16)[name = string("linear_198_cast_fp16")];
            tensor<int32, [4]> var_9690 = const()[name = string("op_9690"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_57_cast_fp16 = reshape(shape = var_9690, x = linear_198_cast_fp16)[name = string("v_57_cast_fp16")];
            tensor<int32, [4]> q_171_perm_0 = const()[name = string("q_171_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_115_perm_0 = const()[name = string("k_115_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_28_perm_0 = const()[name = string("transpose_28_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_171_cast_fp16 = transpose(perm = q_171_perm_0, x = q_169_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 64]> var_9703_cast_fp16 = mul(x = q_171_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9703_cast_fp16")];
            tensor<int32, [4]> x1_113_begin_0 = const()[name = string("x1_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_113_end_0 = const()[name = string("x1_113_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_113_end_mask_0 = const()[name = string("x1_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_113_cast_fp16 = slice_by_index(begin = x1_113_begin_0, end = x1_113_end_0, end_mask = x1_113_end_mask_0, x = q_171_cast_fp16)[name = string("x1_113_cast_fp16")];
            tensor<int32, [4]> x2_113_begin_0 = const()[name = string("x2_113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_113_end_0 = const()[name = string("x2_113_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_113_end_mask_0 = const()[name = string("x2_113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_113_cast_fp16 = slice_by_index(begin = x2_113_begin_0, end = x2_113_end_0, end_mask = x2_113_end_mask_0, x = q_171_cast_fp16)[name = string("x2_113_cast_fp16")];
            fp16 const_343_promoted_to_fp16 = const()[name = string("const_343_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9724_cast_fp16 = mul(x = x2_113_cast_fp16, y = const_343_promoted_to_fp16)[name = string("op_9724_cast_fp16")];
            int32 var_9726 = const()[name = string("op_9726"), val = int32(-1)];
            bool var_9727_interleave_0 = const()[name = string("op_9727_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9727_cast_fp16 = concat(axis = var_9726, interleave = var_9727_interleave_0, values = (var_9724_cast_fp16, x1_113_cast_fp16))[name = string("op_9727_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9730_cast_fp16 = mul(x = var_9727_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9730_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_173_cast_fp16 = add(x = var_9703_cast_fp16, y = var_9730_cast_fp16)[name = string("q_173_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_115_cast_fp16 = transpose(perm = k_115_perm_0, x = k_113_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 16, 16, 64]> var_9735_cast_fp16 = mul(x = k_115_cast_fp16, y = var_1078_cast_fp16)[name = string("op_9735_cast_fp16")];
            tensor<int32, [4]> x1_115_begin_0 = const()[name = string("x1_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_115_end_0 = const()[name = string("x1_115_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_115_end_mask_0 = const()[name = string("x1_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_115_cast_fp16 = slice_by_index(begin = x1_115_begin_0, end = x1_115_end_0, end_mask = x1_115_end_mask_0, x = k_115_cast_fp16)[name = string("x1_115_cast_fp16")];
            tensor<int32, [4]> x2_115_begin_0 = const()[name = string("x2_115_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_115_end_0 = const()[name = string("x2_115_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_115_end_mask_0 = const()[name = string("x2_115_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_115_cast_fp16 = slice_by_index(begin = x2_115_begin_0, end = x2_115_end_0, end_mask = x2_115_end_mask_0, x = k_115_cast_fp16)[name = string("x2_115_cast_fp16")];
            fp16 const_346_promoted_to_fp16 = const()[name = string("const_346_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_9756_cast_fp16 = mul(x = x2_115_cast_fp16, y = const_346_promoted_to_fp16)[name = string("op_9756_cast_fp16")];
            int32 var_9758 = const()[name = string("op_9758"), val = int32(-1)];
            bool var_9759_interleave_0 = const()[name = string("op_9759_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_9759_cast_fp16 = concat(axis = var_9758, interleave = var_9759_interleave_0, values = (var_9756_cast_fp16, x1_115_cast_fp16))[name = string("op_9759_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9762_cast_fp16 = mul(x = var_9759_cast_fp16, y = var_1105_cast_fp16)[name = string("op_9762_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_113_cast_fp16 = add(x = var_9735_cast_fp16, y = var_9762_cast_fp16)[name = string("x_113_cast_fp16")];
            tensor<int32, [4]> var_9769 = const()[name = string("op_9769"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_9781 = const()[name = string("op_9781"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_9770_cast_fp16 = transpose(perm = var_9769, x = x_113_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 1024, 1, 16]> var_9782_cast_fp16 = reshape(shape = var_9781, x = var_9770_cast_fp16)[name = string("op_9782_cast_fp16")];
            tensor<int32, [1]> var_9784_axes_0 = const()[name = string("op_9784_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9784_cast_fp16 = squeeze(axes = var_9784_axes_0, x = var_9782_cast_fp16)[name = string("op_9784_cast_fp16")];
            bool var_9785_transpose_x_0 = const()[name = string("op_9785_transpose_x_0"), val = bool(false)];
            bool var_9785_transpose_y_0 = const()[name = string("op_9785_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9785_cast_fp16 = matmul(transpose_x = var_9785_transpose_x_0, transpose_y = var_9785_transpose_y_0, x = var_9784_cast_fp16, y = block_cache_map)[name = string("op_9785_cast_fp16")];
            tensor<int32, [1]> nk_full_57_axes_0 = const()[name = string("nk_full_57_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_57_cast_fp16 = expand_dims(axes = nk_full_57_axes_0, x = var_9785_cast_fp16)[name = string("nk_full_57_cast_fp16")];
            tensor<int32, [4]> var_9804 = const()[name = string("op_9804"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_28_cast_fp16 = transpose(perm = transpose_28_perm_0, x = v_57_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 1024, 1, 16]> var_9805_cast_fp16 = reshape(shape = var_9804, x = transpose_28_cast_fp16)[name = string("op_9805_cast_fp16")];
            tensor<int32, [1]> var_9807_axes_0 = const()[name = string("op_9807_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_9807_cast_fp16 = squeeze(axes = var_9807_axes_0, x = var_9805_cast_fp16)[name = string("op_9807_cast_fp16")];
            bool var_9808_transpose_x_0 = const()[name = string("op_9808_transpose_x_0"), val = bool(false)];
            bool var_9808_transpose_y_0 = const()[name = string("op_9808_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_9808_cast_fp16 = matmul(transpose_x = var_9808_transpose_x_0, transpose_y = var_9808_transpose_y_0, x = var_9807_cast_fp16, y = block_cache_map)[name = string("op_9808_cast_fp16")];
            tensor<int32, [1]> nv_full_57_axes_0 = const()[name = string("nv_full_57_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_57_cast_fp16 = expand_dims(axes = nv_full_57_axes_0, x = var_9808_cast_fp16)[name = string("nv_full_57_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9814_cast_fp16 = mul(x = lkc_113_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9814_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9815_cast_fp16 = mul(x = nk_full_57_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9815_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_115_cast_fp16 = add(x = var_9814_cast_fp16, y = var_9815_cast_fp16)[name = string("lkc_115_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9821_cast_fp16 = mul(x = lvc_113_cast_fp16, y = var_1189_cast_fp16)[name = string("op_9821_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_9822_cast_fp16 = mul(x = nv_full_57_cast_fp16, y = update_mask_cast_fp16)[name = string("op_9822_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_115_cast_fp16 = add(x = var_9821_cast_fp16, y = var_9822_cast_fp16)[name = string("lvc_115_cast_fp16")];
            tensor<int32, [1]> var_9826_axes_0 = const()[name = string("op_9826_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9826_cast_fp16 = squeeze(axes = var_9826_axes_0, x = lkc_115_cast_fp16)[name = string("op_9826_cast_fp16")];
            tensor<int32, [4]> var_9830 = const()[name = string("op_9830"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_57_cast_fp16 = reshape(shape = var_9830, x = var_9826_cast_fp16)[name = string("kc_57_cast_fp16")];
            tensor<int32, [1]> var_9833_axes_0 = const()[name = string("op_9833_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_9833_cast_fp16 = squeeze(axes = var_9833_axes_0, x = lvc_115_cast_fp16)[name = string("op_9833_cast_fp16")];
            tensor<int32, [4]> var_9837 = const()[name = string("op_9837"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_57_cast_fp16 = reshape(shape = var_9837, x = var_9833_cast_fp16)[name = string("vc_57_cast_fp16")];
            bool var_9839_transpose_x_0 = const()[name = string("op_9839_transpose_x_0"), val = bool(false)];
            bool var_9839_transpose_y_0 = const()[name = string("op_9839_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_9839_cast_fp16 = matmul(transpose_x = var_9839_transpose_x_0, transpose_y = var_9839_transpose_y_0, x = q_173_cast_fp16, y = kc_57_cast_fp16)[name = string("op_9839_cast_fp16")];
            fp16 _inversed_scores_169_y_0_to_fp16 = const()[name = string("_inversed_scores_169_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_169_cast_fp16 = mul(x = var_9839_cast_fp16, y = _inversed_scores_169_y_0_to_fp16)[name = string("_inversed_scores_169_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_171_cast_fp16 = add(x = _inversed_scores_169_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_171_cast_fp16")];
            int32 var_9868 = const()[name = string("op_9868"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_9870_cast_fp16 = softmax(axis = var_9868, x = scores_171_cast_fp16)[name = string("op_9870_cast_fp16")];
            bool attn_out_113_transpose_x_1 = const()[name = string("attn_out_113_transpose_x_1"), val = bool(false)];
            bool attn_out_113_transpose_y_1 = const()[name = string("attn_out_113_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_113_cast_fp16 = matmul(transpose_x = attn_out_113_transpose_x_1, transpose_y = attn_out_113_transpose_y_1, x = var_9870_cast_fp16, y = vc_57_cast_fp16)[name = string("attn_out_113_cast_fp16")];
            tensor<int32, [4]> var_9882_perm_0 = const()[name = string("op_9882_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9884 = const()[name = string("op_9884"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_9882_cast_fp16 = transpose(perm = var_9882_perm_0, x = attn_out_113_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 1024]> input_229_cast_fp16 = reshape(shape = var_9884, x = var_9882_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_28_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_28_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(971139520)))];
            tensor<fp16, [1, 16, 1024]> linear_199_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_28_self_attn_o_proj_weight_to_fp16, x = input_229_cast_fp16)[name = string("linear_199_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_455_cast_fp16 = add(x = hidden_states_447_cast_fp16, y = linear_199_cast_fp16)[name = string("hidden_states_455_cast_fp16")];
            fp16 var_9894_promoted_to_fp16 = const()[name = string("op_9894_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9900_cast_fp16 = pow(x = hidden_states_455_cast_fp16, y = var_9894_promoted_to_fp16)[name = string("op_9900_cast_fp16")];
            tensor<int32, [1]> variance_115_axes_0 = const()[name = string("variance_115_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_115_keep_dims_0 = const()[name = string("variance_115_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_115_cast_fp16 = reduce_mean(axes = variance_115_axes_0, keep_dims = variance_115_keep_dims_0, x = var_9900_cast_fp16)[name = string("variance_115_cast_fp16")];
            fp16 var_9903_to_fp16 = const()[name = string("op_9903_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9904_cast_fp16 = add(x = variance_115_cast_fp16, y = var_9903_to_fp16)[name = string("op_9904_cast_fp16")];
            fp32 var_9905_epsilon_0 = const()[name = string("op_9905_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9905_cast_fp16 = rsqrt(epsilon = var_9905_epsilon_0, x = var_9904_cast_fp16)[name = string("op_9905_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_459_cast_fp16 = mul(x = hidden_states_455_cast_fp16, y = var_9905_cast_fp16)[name = string("hidden_states_459_cast_fp16")];
            tensor<fp16, [1024]> layers_28_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_28_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(973236736)))];
            tensor<fp16, [1, 16, 1024]> input_231_cast_fp16 = mul(x = layers_28_post_attention_layernorm_weight_to_fp16, y = hidden_states_459_cast_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_28_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_28_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(973238848)))];
            tensor<fp16, [1, 16, 4096]> linear_200_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_28_mlp_gate_proj_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_200_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_9918_cast_fp16 = silu(x = linear_200_cast_fp16)[name = string("op_9918_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_28_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_28_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(981627520)))];
            tensor<fp16, [1, 16, 4096]> linear_201_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_28_mlp_up_proj_weight_to_fp16, x = input_231_cast_fp16)[name = string("linear_201_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_235_cast_fp16 = mul(x = var_9918_cast_fp16, y = linear_201_cast_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_28_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_28_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(990016192)))];
            tensor<fp16, [1, 16, 1024]> linear_202_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_28_mlp_down_proj_weight_to_fp16, x = input_235_cast_fp16)[name = string("linear_202_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_463_cast_fp16 = add(x = hidden_states_455_cast_fp16, y = linear_202_cast_fp16)[name = string("hidden_states_463_cast_fp16")];
            tensor<int32, [4]> lkc_117_begin_0 = const()[name = string("lkc_117_begin_0"), val = tensor<int32, [4]>([0, 29696, 0, 0])];
            tensor<int32, [4]> lkc_117_end_0 = const()[name = string("lkc_117_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> lkc_117_end_mask_0 = const()[name = string("lkc_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_117_cast_fp16 = slice_by_index(begin = lkc_117_begin_0, end = lkc_117_end_0, end_mask = lkc_117_end_mask_0, x = key_cache)[name = string("lkc_117_cast_fp16")];
            tensor<int32, [4]> lvc_117_begin_0 = const()[name = string("lvc_117_begin_0"), val = tensor<int32, [4]>([0, 29696, 0, 0])];
            tensor<int32, [4]> lvc_117_end_0 = const()[name = string("lvc_117_end_0"), val = tensor<int32, [4]>([1, 1, 1, 1024])];
            tensor<bool, [4]> lvc_117_end_mask_0 = const()[name = string("lvc_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_117_cast_fp16 = slice_by_index(begin = lvc_117_begin_0, end = lvc_117_end_0, end_mask = lvc_117_end_mask_0, x = value_cache)[name = string("lvc_117_cast_fp16")];
            fp16 var_9949_promoted_to_fp16 = const()[name = string("op_9949_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9955_cast_fp16 = pow(x = hidden_states_463_cast_fp16, y = var_9949_promoted_to_fp16)[name = string("op_9955_cast_fp16")];
            tensor<int32, [1]> variance_117_axes_0 = const()[name = string("variance_117_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_117_keep_dims_0 = const()[name = string("variance_117_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_117_cast_fp16 = reduce_mean(axes = variance_117_axes_0, keep_dims = variance_117_keep_dims_0, x = var_9955_cast_fp16)[name = string("variance_117_cast_fp16")];
            fp16 var_9958_to_fp16 = const()[name = string("op_9958_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_9959_cast_fp16 = add(x = variance_117_cast_fp16, y = var_9958_to_fp16)[name = string("op_9959_cast_fp16")];
            fp32 var_9960_epsilon_0 = const()[name = string("op_9960_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9960_cast_fp16 = rsqrt(epsilon = var_9960_epsilon_0, x = var_9959_cast_fp16)[name = string("op_9960_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_467_cast_fp16 = mul(x = hidden_states_463_cast_fp16, y = var_9960_cast_fp16)[name = string("hidden_states_467_cast_fp16")];
            tensor<fp16, [1024]> layers_29_input_layernorm_weight_to_fp16 = const()[name = string("layers_29_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998404864)))];
            tensor<fp16, [1, 16, 1024]> hidden_117_cast_fp16 = mul(x = layers_29_input_layernorm_weight_to_fp16, y = hidden_states_467_cast_fp16)[name = string("hidden_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_29_self_attn_q_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_q_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(998406976)))];
            tensor<fp16, [1, 16, 1024]> linear_203_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_29_self_attn_q_proj_weight_to_fp16, x = hidden_117_cast_fp16)[name = string("linear_203_cast_fp16")];
            tensor<int32, [4]> var_9984 = const()[name = string("op_9984"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> q_175_cast_fp16 = reshape(shape = var_9984, x = linear_203_cast_fp16)[name = string("q_175_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_29_self_attn_k_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_k_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1000504192)))];
            tensor<fp16, [1, 16, 1024]> linear_204_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_29_self_attn_k_proj_weight_to_fp16, x = hidden_117_cast_fp16)[name = string("linear_204_cast_fp16")];
            tensor<int32, [4]> var_9991 = const()[name = string("op_9991"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> k_117_cast_fp16 = reshape(shape = var_9991, x = linear_204_cast_fp16)[name = string("k_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_29_self_attn_v_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_v_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1002601408)))];
            tensor<fp16, [1, 16, 1024]> linear_205_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_29_self_attn_v_proj_weight_to_fp16, x = hidden_117_cast_fp16)[name = string("linear_205_cast_fp16")];
            tensor<int32, [4]> var_9998 = const()[name = string("op_9998"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<fp16, [1, 16, 16, 64]> v_cast_fp16 = reshape(shape = var_9998, x = linear_205_cast_fp16)[name = string("v_cast_fp16")];
            tensor<int32, [4]> q_177_perm_0 = const()[name = string("q_177_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> k_perm_0 = const()[name = string("k_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [4]> transpose_29_perm_0 = const()[name = string("transpose_29_perm_0"), val = tensor<int32, [4]>([0, 2, 3, 1])];
            tensor<fp16, [1, 16, 16, 64]> q_177_cast_fp16 = transpose(perm = q_177_perm_0, x = q_175_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 64]> var_10011_cast_fp16 = mul(x = q_177_cast_fp16, y = var_1078_cast_fp16)[name = string("op_10011_cast_fp16")];
            tensor<int32, [4]> x1_117_begin_0 = const()[name = string("x1_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_117_end_0 = const()[name = string("x1_117_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_117_end_mask_0 = const()[name = string("x1_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_117_cast_fp16 = slice_by_index(begin = x1_117_begin_0, end = x1_117_end_0, end_mask = x1_117_end_mask_0, x = q_177_cast_fp16)[name = string("x1_117_cast_fp16")];
            tensor<int32, [4]> x2_117_begin_0 = const()[name = string("x2_117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_117_end_0 = const()[name = string("x2_117_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_117_end_mask_0 = const()[name = string("x2_117_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_117_cast_fp16 = slice_by_index(begin = x2_117_begin_0, end = x2_117_end_0, end_mask = x2_117_end_mask_0, x = q_177_cast_fp16)[name = string("x2_117_cast_fp16")];
            fp16 const_355_promoted_to_fp16 = const()[name = string("const_355_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_10032_cast_fp16 = mul(x = x2_117_cast_fp16, y = const_355_promoted_to_fp16)[name = string("op_10032_cast_fp16")];
            int32 var_10034 = const()[name = string("op_10034"), val = int32(-1)];
            bool var_10035_interleave_0 = const()[name = string("op_10035_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_10035_cast_fp16 = concat(axis = var_10034, interleave = var_10035_interleave_0, values = (var_10032_cast_fp16, x1_117_cast_fp16))[name = string("op_10035_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10038_cast_fp16 = mul(x = var_10035_cast_fp16, y = var_1105_cast_fp16)[name = string("op_10038_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> q_cast_fp16 = add(x = var_10011_cast_fp16, y = var_10038_cast_fp16)[name = string("q_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> k_cast_fp16 = transpose(perm = k_perm_0, x = k_117_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 16, 16, 64]> var_10043_cast_fp16 = mul(x = k_cast_fp16, y = var_1078_cast_fp16)[name = string("op_10043_cast_fp16")];
            tensor<int32, [4]> x1_begin_0 = const()[name = string("x1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> x1_end_0 = const()[name = string("x1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 32])];
            tensor<bool, [4]> x1_end_mask_0 = const()[name = string("x1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 32]> x1_cast_fp16 = slice_by_index(begin = x1_begin_0, end = x1_end_0, end_mask = x1_end_mask_0, x = k_cast_fp16)[name = string("x1_cast_fp16")];
            tensor<int32, [4]> x2_begin_0 = const()[name = string("x2_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 32])];
            tensor<int32, [4]> x2_end_0 = const()[name = string("x2_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> x2_end_mask_0 = const()[name = string("x2_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 32]> x2_cast_fp16 = slice_by_index(begin = x2_begin_0, end = x2_end_0, end_mask = x2_end_mask_0, x = k_cast_fp16)[name = string("x2_cast_fp16")];
            fp16 const_358_promoted_to_fp16 = const()[name = string("const_358_promoted_to_fp16"), val = fp16(-0x1p+0)];
            tensor<fp16, [1, 16, 16, 32]> var_10064_cast_fp16 = mul(x = x2_cast_fp16, y = const_358_promoted_to_fp16)[name = string("op_10064_cast_fp16")];
            int32 var_10066 = const()[name = string("op_10066"), val = int32(-1)];
            bool var_10067_interleave_0 = const()[name = string("op_10067_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 64]> var_10067_cast_fp16 = concat(axis = var_10066, interleave = var_10067_interleave_0, values = (var_10064_cast_fp16, x1_cast_fp16))[name = string("op_10067_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_10070_cast_fp16 = mul(x = var_10067_cast_fp16, y = var_1105_cast_fp16)[name = string("op_10070_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> x_117_cast_fp16 = add(x = var_10043_cast_fp16, y = var_10070_cast_fp16)[name = string("x_117_cast_fp16")];
            tensor<int32, [4]> var_10077 = const()[name = string("op_10077"), val = tensor<int32, [4]>([0, 1, 3, 2])];
            tensor<int32, [4]> var_10089 = const()[name = string("op_10089"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> var_10078_cast_fp16 = transpose(perm = var_10077, x = x_117_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 1024, 1, 16]> var_10090_cast_fp16 = reshape(shape = var_10089, x = var_10078_cast_fp16)[name = string("op_10090_cast_fp16")];
            tensor<int32, [1]> var_10092_axes_0 = const()[name = string("op_10092_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_10092_cast_fp16 = squeeze(axes = var_10092_axes_0, x = var_10090_cast_fp16)[name = string("op_10092_cast_fp16")];
            bool var_10093_transpose_x_0 = const()[name = string("op_10093_transpose_x_0"), val = bool(false)];
            bool var_10093_transpose_y_0 = const()[name = string("op_10093_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_10093_cast_fp16 = matmul(transpose_x = var_10093_transpose_x_0, transpose_y = var_10093_transpose_y_0, x = var_10092_cast_fp16, y = block_cache_map)[name = string("op_10093_cast_fp16")];
            tensor<int32, [1]> nk_full_axes_0 = const()[name = string("nk_full_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nk_full_cast_fp16 = expand_dims(axes = nk_full_axes_0, x = var_10093_cast_fp16)[name = string("nk_full_cast_fp16")];
            tensor<int32, [4]> var_10112 = const()[name = string("op_10112"), val = tensor<int32, [4]>([1, 1024, 1, 16])];
            tensor<fp16, [1, 16, 64, 16]> transpose_29_cast_fp16 = transpose(perm = transpose_29_perm_0, x = v_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 1024, 1, 16]> var_10113_cast_fp16 = reshape(shape = var_10112, x = transpose_29_cast_fp16)[name = string("op_10113_cast_fp16")];
            tensor<int32, [1]> var_10115_axes_0 = const()[name = string("op_10115_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 16]> var_10115_cast_fp16 = squeeze(axes = var_10115_axes_0, x = var_10113_cast_fp16)[name = string("op_10115_cast_fp16")];
            bool var_10116_transpose_x_0 = const()[name = string("op_10116_transpose_x_0"), val = bool(false)];
            bool var_10116_transpose_y_0 = const()[name = string("op_10116_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 1024, 1024]> var_10116_cast_fp16 = matmul(transpose_x = var_10116_transpose_x_0, transpose_y = var_10116_transpose_y_0, x = var_10115_cast_fp16, y = block_cache_map)[name = string("op_10116_cast_fp16")];
            tensor<int32, [1]> nv_full_axes_0 = const()[name = string("nv_full_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1, 1024]> nv_full_cast_fp16 = expand_dims(axes = nv_full_axes_0, x = var_10116_cast_fp16)[name = string("nv_full_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_10122_cast_fp16 = mul(x = lkc_117_cast_fp16, y = var_1189_cast_fp16)[name = string("op_10122_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_10123_cast_fp16 = mul(x = nk_full_cast_fp16, y = update_mask_cast_fp16)[name = string("op_10123_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lkc_cast_fp16 = add(x = var_10122_cast_fp16, y = var_10123_cast_fp16)[name = string("lkc_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_10129_cast_fp16 = mul(x = lvc_117_cast_fp16, y = var_1189_cast_fp16)[name = string("op_10129_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> var_10130_cast_fp16 = mul(x = nv_full_cast_fp16, y = update_mask_cast_fp16)[name = string("op_10130_cast_fp16")];
            tensor<fp16, [1, 1024, 1, 1024]> lvc_cast_fp16 = add(x = var_10129_cast_fp16, y = var_10130_cast_fp16)[name = string("lvc_cast_fp16")];
            tensor<int32, [1]> var_10134_axes_0 = const()[name = string("op_10134_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_10134_cast_fp16 = squeeze(axes = var_10134_axes_0, x = lkc_cast_fp16)[name = string("op_10134_cast_fp16")];
            tensor<int32, [4]> var_10138 = const()[name = string("op_10138"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> kc_cast_fp16 = reshape(shape = var_10138, x = var_10134_cast_fp16)[name = string("kc_cast_fp16")];
            tensor<int32, [1]> var_10141_axes_0 = const()[name = string("op_10141_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [1, 1024, 1024]> var_10141_cast_fp16 = squeeze(axes = var_10141_axes_0, x = lvc_cast_fp16)[name = string("op_10141_cast_fp16")];
            tensor<int32, [4]> var_10145 = const()[name = string("op_10145"), val = tensor<int32, [4]>([1, 16, 64, 1024])];
            tensor<fp16, [1, 16, 64, 1024]> vc_cast_fp16 = reshape(shape = var_10145, x = var_10141_cast_fp16)[name = string("vc_cast_fp16")];
            bool var_10147_transpose_x_0 = const()[name = string("op_10147_transpose_x_0"), val = bool(false)];
            bool var_10147_transpose_y_0 = const()[name = string("op_10147_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 1024]> var_10147_cast_fp16 = matmul(transpose_x = var_10147_transpose_x_0, transpose_y = var_10147_transpose_y_0, x = q_cast_fp16, y = kc_cast_fp16)[name = string("op_10147_cast_fp16")];
            fp16 _inversed_scores_175_y_0_to_fp16 = const()[name = string("_inversed_scores_175_y_0_to_fp16"), val = fp16(0x1p-3)];
            tensor<fp16, [1, 16, 16, 1024]> _inversed_scores_175_cast_fp16 = mul(x = var_10147_cast_fp16, y = _inversed_scores_175_y_0_to_fp16)[name = string("_inversed_scores_175_cast_fp16")];
            tensor<fp16, [1, 16, 16, 1024]> scores_177_cast_fp16 = add(x = _inversed_scores_175_cast_fp16, y = var_1226_cast_fp16)[name = string("scores_177_cast_fp16")];
            int32 var_10176 = const()[name = string("op_10176"), val = int32(-1)];
            tensor<fp16, [1, 16, 16, 1024]> var_10178_cast_fp16 = softmax(axis = var_10176, x = scores_177_cast_fp16)[name = string("op_10178_cast_fp16")];
            bool attn_out_117_transpose_x_1 = const()[name = string("attn_out_117_transpose_x_1"), val = bool(false)];
            bool attn_out_117_transpose_y_1 = const()[name = string("attn_out_117_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 64]> attn_out_117_cast_fp16 = matmul(transpose_x = attn_out_117_transpose_x_1, transpose_y = attn_out_117_transpose_y_1, x = var_10178_cast_fp16, y = vc_cast_fp16)[name = string("attn_out_117_cast_fp16")];
            tensor<int32, [4]> var_10190_perm_0 = const()[name = string("op_10190_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_10192 = const()[name = string("op_10192"), val = tensor<int32, [3]>([1, 16, 1024])];
            tensor<fp16, [1, 16, 16, 64]> var_10190_cast_fp16 = transpose(perm = var_10190_perm_0, x = attn_out_117_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 1024]> input_237_cast_fp16 = reshape(shape = var_10192, x = var_10190_cast_fp16)[name = string("input_237_cast_fp16")];
            tensor<fp16, [1024, 1024]> layers_29_self_attn_o_proj_weight_to_fp16 = const()[name = string("layers_29_self_attn_o_proj_weight_to_fp16"), val = tensor<fp16, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1004698624)))];
            tensor<fp16, [1, 16, 1024]> linear_206_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_29_self_attn_o_proj_weight_to_fp16, x = input_237_cast_fp16)[name = string("linear_206_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_471_cast_fp16 = add(x = hidden_states_463_cast_fp16, y = linear_206_cast_fp16)[name = string("hidden_states_471_cast_fp16")];
            fp16 var_10202_promoted_to_fp16 = const()[name = string("op_10202_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10208_cast_fp16 = pow(x = hidden_states_471_cast_fp16, y = var_10202_promoted_to_fp16)[name = string("op_10208_cast_fp16")];
            tensor<int32, [1]> variance_119_axes_0 = const()[name = string("variance_119_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_119_keep_dims_0 = const()[name = string("variance_119_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_119_cast_fp16 = reduce_mean(axes = variance_119_axes_0, keep_dims = variance_119_keep_dims_0, x = var_10208_cast_fp16)[name = string("variance_119_cast_fp16")];
            fp16 var_10211_to_fp16 = const()[name = string("op_10211_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_10212_cast_fp16 = add(x = variance_119_cast_fp16, y = var_10211_to_fp16)[name = string("op_10212_cast_fp16")];
            fp32 var_10213_epsilon_0 = const()[name = string("op_10213_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10213_cast_fp16 = rsqrt(epsilon = var_10213_epsilon_0, x = var_10212_cast_fp16)[name = string("op_10213_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_475_cast_fp16 = mul(x = hidden_states_471_cast_fp16, y = var_10213_cast_fp16)[name = string("hidden_states_475_cast_fp16")];
            tensor<fp16, [1024]> layers_29_post_attention_layernorm_weight_to_fp16 = const()[name = string("layers_29_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1006795840)))];
            tensor<fp16, [1, 16, 1024]> input_239_cast_fp16 = mul(x = layers_29_post_attention_layernorm_weight_to_fp16, y = hidden_states_475_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_29_mlp_gate_proj_weight_to_fp16 = const()[name = string("layers_29_mlp_gate_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1006797952)))];
            tensor<fp16, [1, 16, 4096]> linear_207_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_29_mlp_gate_proj_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_207_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> var_10226_cast_fp16 = silu(x = linear_207_cast_fp16)[name = string("op_10226_cast_fp16")];
            tensor<fp16, [4096, 1024]> layers_29_mlp_up_proj_weight_to_fp16 = const()[name = string("layers_29_mlp_up_proj_weight_to_fp16"), val = tensor<fp16, [4096, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1015186624)))];
            tensor<fp16, [1, 16, 4096]> linear_208_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = layers_29_mlp_up_proj_weight_to_fp16, x = input_239_cast_fp16)[name = string("linear_208_cast_fp16")];
            tensor<fp16, [1, 16, 4096]> input_243_cast_fp16 = mul(x = var_10226_cast_fp16, y = linear_208_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 4096]> layers_29_mlp_down_proj_weight_to_fp16 = const()[name = string("layers_29_mlp_down_proj_weight_to_fp16"), val = tensor<fp16, [1024, 4096]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1023575296)))];
            tensor<fp16, [1, 16, 1024]> linear_209_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = layers_29_mlp_down_proj_weight_to_fp16, x = input_243_cast_fp16)[name = string("linear_209_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_479_cast_fp16 = add(x = hidden_states_471_cast_fp16, y = linear_209_cast_fp16)[name = string("hidden_states_479_cast_fp16")];
            fp16 var_10237_promoted_to_fp16 = const()[name = string("op_10237_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_10243_cast_fp16 = pow(x = hidden_states_479_cast_fp16, y = var_10237_promoted_to_fp16)[name = string("op_10243_cast_fp16")];
            tensor<int32, [1]> variance_axes_0 = const()[name = string("variance_axes_0"), val = tensor<int32, [1]>([-1])];
            bool variance_keep_dims_0 = const()[name = string("variance_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> variance_cast_fp16 = reduce_mean(axes = variance_axes_0, keep_dims = variance_keep_dims_0, x = var_10243_cast_fp16)[name = string("variance_cast_fp16")];
            fp16 var_10246_to_fp16 = const()[name = string("op_10246_to_fp16"), val = fp16(0x1.5p-17)];
            tensor<fp16, [1, 16, 1]> var_10247_cast_fp16 = add(x = variance_cast_fp16, y = var_10246_to_fp16)[name = string("op_10247_cast_fp16")];
            fp32 var_10248_epsilon_0 = const()[name = string("op_10248_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_10248_cast_fp16 = rsqrt(epsilon = var_10248_epsilon_0, x = var_10247_cast_fp16)[name = string("op_10248_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> hidden_states_483_cast_fp16 = mul(x = hidden_states_479_cast_fp16, y = var_10248_cast_fp16)[name = string("hidden_states_483_cast_fp16")];
            tensor<fp16, [1024]> norm_weight_to_fp16 = const()[name = string("norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031963968)))];
            tensor<fp16, [1, 16, 1024]> block_hidden = mul(x = norm_weight_to_fp16, y = hidden_states_483_cast_fp16)[name = string("hidden_cast_fp16")];
            int32 var_10253 = const()[name = string("op_10253"), val = int32(1)];
            bool var_10254_interleave_0 = const()[name = string("op_10254_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 30720, 1, 1024]> new_key_cache = concat(axis = var_10253, interleave = var_10254_interleave_0, values = (lkc_3_cast_fp16, lkc_7_cast_fp16, lkc_11_cast_fp16, lkc_15_cast_fp16, lkc_19_cast_fp16, lkc_23_cast_fp16, lkc_27_cast_fp16, lkc_31_cast_fp16, lkc_35_cast_fp16, lkc_39_cast_fp16, lkc_43_cast_fp16, lkc_47_cast_fp16, lkc_51_cast_fp16, lkc_55_cast_fp16, lkc_59_cast_fp16, lkc_63_cast_fp16, lkc_67_cast_fp16, lkc_71_cast_fp16, lkc_75_cast_fp16, lkc_79_cast_fp16, lkc_83_cast_fp16, lkc_87_cast_fp16, lkc_91_cast_fp16, lkc_95_cast_fp16, lkc_99_cast_fp16, lkc_103_cast_fp16, lkc_107_cast_fp16, lkc_111_cast_fp16, lkc_115_cast_fp16, lkc_cast_fp16))[name = string("op_10254_cast_fp16")];
            int32 var_10256 = const()[name = string("op_10256"), val = int32(1)];
            bool var_10257_interleave_0 = const()[name = string("op_10257_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 30720, 1, 1024]> new_value_cache = concat(axis = var_10256, interleave = var_10257_interleave_0, values = (lvc_3_cast_fp16, lvc_7_cast_fp16, lvc_11_cast_fp16, lvc_15_cast_fp16, lvc_19_cast_fp16, lvc_23_cast_fp16, lvc_27_cast_fp16, lvc_31_cast_fp16, lvc_35_cast_fp16, lvc_39_cast_fp16, lvc_43_cast_fp16, lvc_47_cast_fp16, lvc_51_cast_fp16, lvc_55_cast_fp16, lvc_59_cast_fp16, lvc_63_cast_fp16, lvc_67_cast_fp16, lvc_71_cast_fp16, lvc_75_cast_fp16, lvc_79_cast_fp16, lvc_83_cast_fp16, lvc_87_cast_fp16, lvc_91_cast_fp16, lvc_95_cast_fp16, lvc_99_cast_fp16, lvc_103_cast_fp16, lvc_107_cast_fp16, lvc_111_cast_fp16, lvc_115_cast_fp16, lvc_cast_fp16))[name = string("op_10257_cast_fp16")];
            tensor<int32, [3]> var_10267_begin_0 = const()[name = string("op_10267_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_10267_end_0 = const()[name = string("op_10267_end_0"), val = tensor<int32, [3]>([1, 15, 1024])];
            tensor<bool, [3]> var_10267_end_mask_0 = const()[name = string("op_10267_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<fp16, [1, 15, 1024]> var_10267_cast_fp16 = slice_by_index(begin = var_10267_begin_0, end = var_10267_end_0, end_mask = var_10267_end_mask_0, x = block_hidden)[name = string("op_10267_cast_fp16")];
            int32 var_10269 = const()[name = string("op_10269"), val = int32(1)];
            bool input_interleave_0 = const()[name = string("input_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 1024]> input_cast_fp16 = concat(axis = var_10269, interleave = input_interleave_0, values = (shift_ctx, var_10267_cast_fp16))[name = string("input_cast_fp16")];
            tensor<fp16, [8194, 1024]> speech_head_weight_to_fp16 = const()[name = string("speech_head_weight_to_fp16"), val = tensor<fp16, [8194, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1031966080)))];
            tensor<fp16, [8194]> linear_210_bias_0_to_fp16 = const()[name = string("linear_210_bias_0_to_fp16"), val = tensor<fp16, [8194]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1048747456)))];
            tensor<fp16, [1, 16, 8194]> logits = linear(bias = linear_210_bias_0_to_fp16, weight = speech_head_weight_to_fp16, x = input_cast_fp16)[name = string("linear_210_cast_fp16")];
        } -> (logits, block_hidden, new_key_cache, new_value_cache);
}