program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3600.16.1"}, {"coremlc-version", "3600.25.1"}})]
{
    func window2048<ios18>(tensor<int32, [1]> answerRow, tensor<fp16, [1, 1, 1, 2048]> cacheMask, tensor<int32, [1, 16]> inputIds, state<tensor<fp16, [28, 8, 2064, 128]>> keyCache, tensor<int32, [16]> positions, state<tensor<fp16, [28, 8, 2064, 128]>> valueCache) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> model_embed_tokens_weight_to_fp16 = const()[name = string("model_embed_tokens_weight_to_fp16"), val = tensor<fp16, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = inputIds, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0 = add(x = inputIds, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = inputIds, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 16]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 x_1_cast_fp16_axis_0 = const()[name = string("x_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 16, 1024]> x_1_cast_fp16 = gather(axis = x_1_cast_fp16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = model_embed_tokens_weight_to_fp16)[name = string("x_1_cast_fp16")];
            int32 var_1504 = const()[name = string("op_1504"), val = int32(0)];
            int32 var_1505_batch_dims_0 = const()[name = string("op_1505_batch_dims_0"), val = int32(0)];
            bool var_1505_validate_indices_0 = const()[name = string("op_1505_validate_indices_0"), val = bool(false)];
            tensor<fp16, [2064, 64]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2064, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752806016))), scale = tensor<fp16, [2064, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752938176))))[name = string("cos_to_fp16_quantized")];
            string positions_to_uint16_dtype_0 = const()[name = string("positions_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [16]> positions_to_uint16 = cast(dtype = positions_to_uint16_dtype_0, x = positions)[name = string("cast_1")];
            tensor<fp16, [16, 64]> var_1505_cast_fp16_cast_uint16 = gather(axis = var_1504, batch_dims = var_1505_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1505_validate_indices_0, x = cos_to_fp16_quantized)[name = string("op_1505_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> cos_1_cast_fp16 = reshape(shape = var_1510, x = var_1505_cast_fp16_cast_uint16)[name = string("cos_1_cast_fp16")];
            int32 var_1512 = const()[name = string("op_1512"), val = int32(0)];
            int32 var_1513_batch_dims_0 = const()[name = string("op_1513_batch_dims_0"), val = int32(0)];
            bool var_1513_validate_indices_0 = const()[name = string("op_1513_validate_indices_0"), val = bool(false)];
            tensor<fp16, [2064, 64]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2064, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752942400))), scale = tensor<fp16, [2064, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753074560))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [16, 64]> var_1513_cast_fp16_cast_uint16 = gather(axis = var_1512, batch_dims = var_1513_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1513_validate_indices_0, x = sin_to_fp16_quantized)[name = string("op_1513_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1518 = const()[name = string("op_1518"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> sin_1_cast_fp16 = reshape(shape = var_1518, x = var_1513_cast_fp16_cast_uint16)[name = string("sin_1_cast_fp16")];
            fp16 var_1539_promoted_to_fp16 = const()[name = string("op_1539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = pow(x = x_1_cast_fp16, y = var_1539_promoted_to_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1541_cast_fp16)[name = string("op_1543_cast_fp16")];
            fp16 var_1544_to_fp16 = const()[name = string("op_1544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1545_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1544_to_fp16)[name = string("op_1545_cast_fp16")];
            fp32 var_1546_epsilon_0 = const()[name = string("op_1546_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1546_cast_fp16 = rsqrt(epsilon = var_1546_epsilon_0, x = var_1545_cast_fp16)[name = string("op_1546_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1547_cast_fp16 = mul(x = x_1_cast_fp16, y = var_1546_cast_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311235072)))];
            tensor<fp16, [1, 16, 1024]> input_1_cast_fp16 = mul(x = var_1547_cast_fp16, y = model_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311237184))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313334400))))[name = string("model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313338560)))];
            tensor<fp16, [1, 16, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_1556 = const()[name = string("op_1556"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1557_cast_fp16 = reshape(shape = var_1556, x = linear_0_cast_fp16)[name = string("op_1557_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1565_promoted_to_fp16 = const()[name = string("op_1565_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_1557_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1567_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1565_promoted_to_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [1]> var_1569_axes_0 = const()[name = string("op_1569_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1569_keep_dims_0 = const()[name = string("op_1569_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1569_cast_fp16 = reduce_mean(axes = var_1569_axes_0, keep_dims = var_1569_keep_dims_0, x = var_1567_cast_fp16)[name = string("op_1569_cast_fp16")];
            fp16 var_1570_to_fp16 = const()[name = string("op_1570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1571_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_to_fp16)[name = string("op_1571_cast_fp16")];
            fp32 var_1572_epsilon_0 = const()[name = string("op_1572_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1572_cast_fp16 = rsqrt(epsilon = var_1572_epsilon_0, x = var_1571_cast_fp16)[name = string("op_1572_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1573_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1572_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313342720)))];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = mul(x = var_1573_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313343040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314391680))))[name = string("model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314393792)))];
            tensor<fp16, [1, 16, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = string("op_1582"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1583_cast_fp16 = reshape(shape = var_1582, x = linear_1_cast_fp16)[name = string("op_1583_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1591_promoted_to_fp16 = const()[name = string("op_1591_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1583_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1593_cast_fp16 = pow(x = x_5_cast_fp16, y = var_1591_promoted_to_fp16)[name = string("op_1593_cast_fp16")];
            tensor<int32, [1]> var_1595_axes_0 = const()[name = string("op_1595_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1595_keep_dims_0 = const()[name = string("op_1595_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1595_cast_fp16 = reduce_mean(axes = var_1595_axes_0, keep_dims = var_1595_keep_dims_0, x = var_1593_cast_fp16)[name = string("op_1595_cast_fp16")];
            fp16 var_1596_to_fp16 = const()[name = string("op_1596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1597_cast_fp16 = add(x = var_1595_cast_fp16, y = var_1596_to_fp16)[name = string("op_1597_cast_fp16")];
            fp32 var_1598_epsilon_0 = const()[name = string("op_1598_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1598_cast_fp16 = rsqrt(epsilon = var_1598_epsilon_0, x = var_1597_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1599_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1598_cast_fp16)[name = string("op_1599_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314395904)))];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = mul(x = var_1599_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314396224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315444864))))[name = string("model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1608 = const()[name = string("op_1608"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1609_cast_fp16 = reshape(shape = var_1608, x = linear_2_cast_fp16)[name = string("op_1609_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_1_begin_0 = const()[name = string("first_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_1_end_0 = const()[name = string("first_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_1_end_mask_0 = const()[name = string("first_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_1_cast_fp16 = slice_by_index(begin = first_1_begin_0, end = first_1_end_0, end_mask = first_1_end_mask_0, x = x_7_cast_fp16)[name = string("first_1_cast_fp16")];
            tensor<int32, [4]> second_1_begin_0 = const()[name = string("second_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_1_end_0 = const()[name = string("second_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_1_end_mask_0 = const()[name = string("second_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_1_cast_fp16 = slice_by_index(begin = second_1_begin_0, end = second_1_end_0, end_mask = second_1_end_mask_0, x = x_7_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1623_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1624_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1626_cast_fp16 = sub(x = var_1623_cast_fp16, y = var_1624_cast_fp16)[name = string("op_1626_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1627_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1628_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1630_cast_fp16 = add(x = var_1627_cast_fp16, y = var_1628_cast_fp16)[name = string("op_1630_cast_fp16")];
            int32 var_1632 = const()[name = string("op_1632"), val = int32(-1)];
            bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_1_cast_fp16 = concat(axis = var_1632, interleave = q_1_interleave_0, values = (var_1626_cast_fp16, var_1630_cast_fp16))[name = string("q_1_cast_fp16")];
            tensor<int32, [4]> first_3_begin_0 = const()[name = string("first_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_3_end_0 = const()[name = string("first_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_3_end_mask_0 = const()[name = string("first_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_3_cast_fp16 = slice_by_index(begin = first_3_begin_0, end = first_3_end_0, end_mask = first_3_end_mask_0, x = x_9_cast_fp16)[name = string("first_3_cast_fp16")];
            tensor<int32, [4]> second_3_begin_0 = const()[name = string("second_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_3_end_0 = const()[name = string("second_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_3_end_mask_0 = const()[name = string("second_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_3_cast_fp16 = slice_by_index(begin = second_3_begin_0, end = second_3_end_0, end_mask = second_3_end_mask_0, x = x_9_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1644_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1645_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1647_cast_fp16 = sub(x = var_1644_cast_fp16, y = var_1645_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1648_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1649_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1651_cast_fp16 = add(x = var_1648_cast_fp16, y = var_1649_cast_fp16)[name = string("op_1651_cast_fp16")];
            int32 var_1653 = const()[name = string("op_1653"), val = int32(-1)];
            bool k_1_interleave_0 = const()[name = string("k_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_1_cast_fp16 = concat(axis = var_1653, interleave = k_1_interleave_0, values = (var_1647_cast_fp16, var_1651_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1664_begin_0 = const()[name = string("op_1664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1664_end_0 = const()[name = string("op_1664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1664_end_mask_0 = const()[name = string("op_1664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1664_stride_0 = const()[name = string("op_1664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, stride = var_1664_stride_0, x = q_1_cast_fp16)[name = string("op_1664_cast_fp16")];
            tensor<int32, [4]> var_1674_begin_0 = const()[name = string("op_1674_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1674_end_0 = const()[name = string("op_1674_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1674_end_mask_0 = const()[name = string("op_1674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1674_stride_0 = const()[name = string("op_1674_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1674_cast_fp16 = slice_by_index(begin = var_1674_begin_0, end = var_1674_end_0, end_mask = var_1674_end_mask_0, stride = var_1674_stride_0, x = q_1_cast_fp16)[name = string("op_1674_cast_fp16")];
            int32 var_1676 = const()[name = string("op_1676"), val = int32(2)];
            bool stacked_1_interleave_0 = const()[name = string("stacked_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_1_cast_fp16 = concat(axis = var_1676, interleave = stacked_1_interleave_0, values = (var_1664_cast_fp16, var_1674_cast_fp16))[name = string("stacked_1_cast_fp16")];
            tensor<fp16, [28, 8, 2064, 128]> read_state_0 = read_state(input = keyCache)[name = string("read_state_0")];
            tensor<int32, [4]> var_1682_begin_0 = const()[name = string("op_1682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1682_end_0 = const()[name = string("op_1682_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_1682_end_mask_0 = const()[name = string("op_1682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, x = read_state_0)[name = string("op_1682_cast_fp16")];
            tensor<int32, [4]> var_1692_begin_0 = const()[name = string("op_1692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1692_end_0 = const()[name = string("op_1692_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_1692_end_mask_0 = const()[name = string("op_1692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_1692_cast_fp16 = slice_by_index(begin = var_1692_begin_0, end = var_1692_end_0, end_mask = var_1692_end_mask_0, x = var_1682_cast_fp16)[name = string("op_1692_cast_fp16")];
            bool var_1701_transpose_x_1 = const()[name = string("op_1701_transpose_x_1"), val = bool(false)];
            bool var_1701_transpose_y_1 = const()[name = string("op_1701_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_1701_cast_fp16 = matmul(transpose_x = var_1701_transpose_x_1, transpose_y = var_1701_transpose_y_1, x = var_1692_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> var_1704_perm_0 = const()[name = string("op_1704_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_1704_cast_fp16 = transpose(perm = var_1704_perm_0, x = var_1701_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 2048]> var_1706_cast_fp16 = mul(x = var_1704_cast_fp16, y = var_1705_to_fp16)[name = string("op_1706_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_1_cast_fp16 = add(x = var_1706_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1712_transpose_x_1 = const()[name = string("op_1712_transpose_x_1"), val = bool(false)];
            bool var_1712_transpose_y_1 = const()[name = string("op_1712_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1712_cast_fp16 = matmul(transpose_x = var_1712_transpose_x_1, transpose_y = var_1712_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1712_cast_fp16")];
            fp16 var_1713_to_fp16 = const()[name = string("op_1713_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1714_cast_fp16 = mul(x = var_1712_cast_fp16, y = var_1713_to_fp16)[name = string("op_1714_cast_fp16")];
            tensor<fp16, [32, 16]> later_rows_mask_to_fp16 = const()[name = string("later_rows_mask_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315446976)))];
            tensor<fp16, [1, 8, 32, 16]> call_weights_1_cast_fp16 = add(x = var_1714_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_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, 8, 32, 2064]> var_1719_cast_fp16 = concat(axis = var_1718, interleave = var_1719_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1719_cast_fp16")];
            int32 var_1720 = const()[name = string("op_1720"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_1_cast_fp16 = softmax(axis = var_1720, x = var_1719_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1727_begin_0 = const()[name = string("op_1727_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1727_end_0 = const()[name = string("op_1727_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_1727_end_mask_0 = const()[name = string("op_1727_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_1727_cast_fp16 = slice_by_index(begin = var_1727_begin_0, end = var_1727_end_0, end_mask = var_1727_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [28, 8, 2064, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = read_state_1)[name = string("op_1732_cast_fp16")];
            tensor<int32, [4]> var_1742_begin_0 = const()[name = string("op_1742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1742_end_0 = const()[name = string("op_1742_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_1742_end_mask_0 = const()[name = string("op_1742_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_1742_cast_fp16 = slice_by_index(begin = var_1742_begin_0, end = var_1742_end_0, end_mask = var_1742_end_mask_0, x = var_1732_cast_fp16)[name = string("op_1742_cast_fp16")];
            bool var_1748_transpose_x_0 = const()[name = string("op_1748_transpose_x_0"), val = bool(false)];
            bool var_1748_transpose_y_0 = const()[name = string("op_1748_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1748_cast_fp16 = matmul(transpose_x = var_1748_transpose_x_0, transpose_y = var_1748_transpose_y_0, x = var_1727_cast_fp16, y = var_1742_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1753_cast_fp16 = slice_by_index(begin = var_1753_begin_0, end = var_1753_end_0, end_mask = var_1753_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1753_cast_fp16")];
            bool var_1754_transpose_x_0 = const()[name = string("op_1754_transpose_x_0"), val = bool(false)];
            bool var_1754_transpose_y_0 = const()[name = string("op_1754_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1609_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1754_cast_fp16 = matmul(transpose_x = var_1754_transpose_x_0, transpose_y = var_1754_transpose_y_0, x = var_1753_cast_fp16, y = v_1_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1754_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1771_begin_0 = const()[name = string("op_1771_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1771_end_0 = const()[name = string("op_1771_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1771_end_mask_0 = const()[name = string("op_1771_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1771_cast_fp16 = slice_by_index(begin = var_1771_begin_0, end = var_1771_end_0, end_mask = var_1771_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1771_cast_fp16")];
            tensor<int32, [4]> var_1786_begin_0 = const()[name = string("op_1786_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1786_end_0 = const()[name = string("op_1786_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1786_end_mask_0 = const()[name = string("op_1786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1786_cast_fp16 = slice_by_index(begin = var_1786_begin_0, end = var_1786_end_0, end_mask = var_1786_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1788 = const()[name = string("op_1788"), val = int32(1)];
            bool mixed_3_interleave_0 = const()[name = string("mixed_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_3_cast_fp16 = concat(axis = var_1788, interleave = mixed_3_interleave_0, values = (var_1771_cast_fp16, var_1786_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1792_perm_0 = const()[name = string("op_1792_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1792_cast_fp16 = transpose(perm = var_1792_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1796, x = var_1792_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315448064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317545280))))[name = string("model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_11_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1807_promoted_to_fp16 = const()[name = string("op_1807_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1809_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1807_promoted_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1811_axes_0 = const()[name = string("op_1811_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1811_keep_dims_0 = const()[name = string("op_1811_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1811_cast_fp16 = reduce_mean(axes = var_1811_axes_0, keep_dims = var_1811_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1811_cast_fp16")];
            fp16 var_1812_to_fp16 = const()[name = string("op_1812_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1813_cast_fp16 = add(x = var_1811_cast_fp16, y = var_1812_to_fp16)[name = string("op_1813_cast_fp16")];
            fp32 var_1814_epsilon_0 = const()[name = string("op_1814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1814_cast_fp16 = rsqrt(epsilon = var_1814_epsilon_0, x = var_1813_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1815_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1814_cast_fp16)[name = string("op_1815_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317547392)))];
            tensor<fp16, [1, 16, 1024]> input_5_cast_fp16 = mul(x = var_1815_cast_fp16, y = model_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317549504))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320695296))))[name = string("model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320701504)))];
            tensor<fp16, [1, 16, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_1820_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320707712))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323853504))))[name = string("model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_9_cast_fp16 = mul(x = var_1820_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323859712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327005504))))[name = string("model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_6_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1834_promoted_to_fp16 = const()[name = string("op_1834_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1836_cast_fp16 = pow(x = x_13_cast_fp16, y = var_1834_promoted_to_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [1]> var_1838_axes_0 = const()[name = string("op_1838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1838_keep_dims_0 = const()[name = string("op_1838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = reduce_mean(axes = var_1838_axes_0, keep_dims = var_1838_keep_dims_0, x = var_1836_cast_fp16)[name = string("op_1838_cast_fp16")];
            fp16 var_1839_to_fp16 = const()[name = string("op_1839_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1840_cast_fp16 = add(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            fp32 var_1841_epsilon_0 = const()[name = string("op_1841_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = rsqrt(epsilon = var_1841_epsilon_0, x = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1842_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327007616)))];
            tensor<fp16, [1, 16, 1024]> input_11_cast_fp16 = mul(x = var_1842_cast_fp16, y = model_layers_1_input_layernorm_weight_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327009728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329106944))))[name = string("model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_1851 = const()[name = string("op_1851"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1852_cast_fp16 = reshape(shape = var_1851, x = linear_7_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x_15_perm_0 = const()[name = string("x_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_15_cast_fp16 = transpose(perm = x_15_perm_0, x = var_1852_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1862_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [1]> var_1864_axes_0 = const()[name = string("op_1864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1864_keep_dims_0 = const()[name = string("op_1864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1864_cast_fp16 = reduce_mean(axes = var_1864_axes_0, keep_dims = var_1864_keep_dims_0, x = var_1862_cast_fp16)[name = string("op_1864_cast_fp16")];
            fp16 var_1865_to_fp16 = const()[name = string("op_1865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1866_cast_fp16 = add(x = var_1864_cast_fp16, y = var_1865_to_fp16)[name = string("op_1866_cast_fp16")];
            fp32 var_1867_epsilon_0 = const()[name = string("op_1867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1867_cast_fp16 = rsqrt(epsilon = var_1867_epsilon_0, x = var_1866_cast_fp16)[name = string("op_1867_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1868_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1867_cast_fp16)[name = string("op_1868_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111104)))];
            tensor<fp16, [1, 16, 16, 128]> x_19_cast_fp16 = mul(x = var_1868_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330160064))))[name = string("model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1877 = const()[name = string("op_1877"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1878_cast_fp16 = reshape(shape = var_1877, x = linear_8_cast_fp16)[name = string("op_1878_cast_fp16")];
            tensor<int32, [4]> x_17_perm_0 = const()[name = string("x_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1886_promoted_to_fp16 = const()[name = string("op_1886_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = var_1878_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1888_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1886_promoted_to_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [1]> var_1890_axes_0 = const()[name = string("op_1890_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1890_keep_dims_0 = const()[name = string("op_1890_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1890_cast_fp16 = reduce_mean(axes = var_1890_axes_0, keep_dims = var_1890_keep_dims_0, x = var_1888_cast_fp16)[name = string("op_1890_cast_fp16")];
            fp16 var_1891_to_fp16 = const()[name = string("op_1891_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1892_cast_fp16 = add(x = var_1890_cast_fp16, y = var_1891_to_fp16)[name = string("op_1892_cast_fp16")];
            fp32 var_1893_epsilon_0 = const()[name = string("op_1893_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1893_cast_fp16 = rsqrt(epsilon = var_1893_epsilon_0, x = var_1892_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1894_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1893_cast_fp16)[name = string("op_1894_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162176)))];
            tensor<fp16, [1, 8, 16, 128]> x_21_cast_fp16 = mul(x = var_1894_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331211136))))[name = string("model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1903 = const()[name = string("op_1903"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1904_cast_fp16 = reshape(shape = var_1903, x = linear_9_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_5_begin_0 = const()[name = string("first_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_5_end_0 = const()[name = string("first_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_5_end_mask_0 = const()[name = string("first_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_5_cast_fp16 = slice_by_index(begin = first_5_begin_0, end = first_5_end_0, end_mask = first_5_end_mask_0, x = x_19_cast_fp16)[name = string("first_5_cast_fp16")];
            tensor<int32, [4]> second_5_begin_0 = const()[name = string("second_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_5_end_0 = const()[name = string("second_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_5_end_mask_0 = const()[name = string("second_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_5_cast_fp16 = slice_by_index(begin = second_5_begin_0, end = second_5_end_0, end_mask = second_5_end_mask_0, x = x_19_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1918_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1919_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1921_cast_fp16 = sub(x = var_1918_cast_fp16, y = var_1919_cast_fp16)[name = string("op_1921_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1922_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1923_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1925_cast_fp16 = add(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1925_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(-1)];
            bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_3_cast_fp16 = concat(axis = var_1927, interleave = q_3_interleave_0, values = (var_1921_cast_fp16, var_1925_cast_fp16))[name = string("q_3_cast_fp16")];
            tensor<int32, [4]> first_7_begin_0 = const()[name = string("first_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_7_end_0 = const()[name = string("first_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_7_end_mask_0 = const()[name = string("first_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_7_cast_fp16 = slice_by_index(begin = first_7_begin_0, end = first_7_end_0, end_mask = first_7_end_mask_0, x = x_21_cast_fp16)[name = string("first_7_cast_fp16")];
            tensor<int32, [4]> second_7_begin_0 = const()[name = string("second_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_7_end_0 = const()[name = string("second_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_7_end_mask_0 = const()[name = string("second_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_7_cast_fp16 = slice_by_index(begin = second_7_begin_0, end = second_7_end_0, end_mask = second_7_end_mask_0, x = x_21_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1939_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1940_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1942_cast_fp16 = sub(x = var_1939_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1943_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1943_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1944_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1944_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1946_cast_fp16 = add(x = var_1943_cast_fp16, y = var_1944_cast_fp16)[name = string("op_1946_cast_fp16")];
            int32 var_1948 = const()[name = string("op_1948"), val = int32(-1)];
            bool k_3_interleave_0 = const()[name = string("k_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_3_cast_fp16 = concat(axis = var_1948, interleave = k_3_interleave_0, values = (var_1942_cast_fp16, var_1946_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_1959_begin_0 = const()[name = string("op_1959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1959_end_0 = const()[name = string("op_1959_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1959_end_mask_0 = const()[name = string("op_1959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1959_stride_0 = const()[name = string("op_1959_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, stride = var_1959_stride_0, x = q_3_cast_fp16)[name = string("op_1959_cast_fp16")];
            tensor<int32, [4]> var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1969_stride_0 = const()[name = string("op_1969_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, stride = var_1969_stride_0, x = q_3_cast_fp16)[name = string("op_1969_cast_fp16")];
            int32 var_1971 = const()[name = string("op_1971"), val = int32(2)];
            bool stacked_3_interleave_0 = const()[name = string("stacked_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_3_cast_fp16 = concat(axis = var_1971, interleave = stacked_3_interleave_0, values = (var_1959_cast_fp16, var_1969_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_1977_begin_0 = const()[name = string("op_1977_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1977_end_0 = const()[name = string("op_1977_end_0"), val = tensor<int32, [4]>([2, 8, 2064, 128])];
            tensor<bool, [4]> var_1977_end_mask_0 = const()[name = string("op_1977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, x = read_state_0)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> var_1987_begin_0 = const()[name = string("op_1987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1987_end_0 = const()[name = string("op_1987_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_1987_end_mask_0 = const()[name = string("op_1987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, x = var_1977_cast_fp16)[name = string("op_1987_cast_fp16")];
            bool var_1996_transpose_x_1 = const()[name = string("op_1996_transpose_x_1"), val = bool(false)];
            bool var_1996_transpose_y_1 = const()[name = string("op_1996_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_1996_cast_fp16 = matmul(transpose_x = var_1996_transpose_x_1, transpose_y = var_1996_transpose_y_1, x = var_1987_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_1996_cast_fp16")];
            tensor<int32, [4]> var_1999_perm_0 = const()[name = string("op_1999_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_1999_cast_fp16 = transpose(perm = var_1999_perm_0, x = var_1996_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 2048]> var_2001_cast_fp16 = mul(x = var_1999_cast_fp16, y = var_2000_to_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_3_cast_fp16 = add(x = var_2001_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2007_transpose_x_1 = const()[name = string("op_2007_transpose_x_1"), val = bool(false)];
            bool var_2007_transpose_y_1 = const()[name = string("op_2007_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2007_cast_fp16 = matmul(transpose_x = var_2007_transpose_x_1, transpose_y = var_2007_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2007_cast_fp16")];
            fp16 var_2008_to_fp16 = const()[name = string("op_2008_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2009_cast_fp16 = mul(x = var_2007_cast_fp16, y = var_2008_to_fp16)[name = string("op_2009_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2009_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2013 = const()[name = string("op_2013"), val = int32(-1)];
            bool var_2014_interleave_0 = const()[name = string("op_2014_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2014_cast_fp16 = concat(axis = var_2013, interleave = var_2014_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2014_cast_fp16")];
            int32 var_2015 = const()[name = string("op_2015"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_3_cast_fp16 = softmax(axis = var_2015, x = var_2014_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2022_begin_0 = const()[name = string("op_2022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2022_end_0 = const()[name = string("op_2022_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2022_end_mask_0 = const()[name = string("op_2022_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2022_cast_fp16 = slice_by_index(begin = var_2022_begin_0, end = var_2022_end_0, end_mask = var_2022_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<int32, [4]> var_2027_begin_0 = const()[name = string("op_2027_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2027_end_0 = const()[name = string("op_2027_end_0"), val = tensor<int32, [4]>([2, 8, 2064, 128])];
            tensor<bool, [4]> var_2027_end_mask_0 = const()[name = string("op_2027_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2027_cast_fp16 = slice_by_index(begin = var_2027_begin_0, end = var_2027_end_0, end_mask = var_2027_end_mask_0, x = read_state_1)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2037_begin_0 = const()[name = string("op_2037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2037_end_0 = const()[name = string("op_2037_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2037_end_mask_0 = const()[name = string("op_2037_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2037_cast_fp16 = slice_by_index(begin = var_2037_begin_0, end = var_2037_end_0, end_mask = var_2037_end_mask_0, x = var_2027_cast_fp16)[name = string("op_2037_cast_fp16")];
            bool var_2043_transpose_x_0 = const()[name = string("op_2043_transpose_x_0"), val = bool(false)];
            bool var_2043_transpose_y_0 = const()[name = string("op_2043_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2043_cast_fp16 = matmul(transpose_x = var_2043_transpose_x_0, transpose_y = var_2043_transpose_y_0, x = var_2022_cast_fp16, y = var_2037_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [4]> var_2048_begin_0 = const()[name = string("op_2048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2048_end_0 = const()[name = string("op_2048_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2048_end_mask_0 = const()[name = string("op_2048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2048_cast_fp16")];
            bool var_2049_transpose_x_0 = const()[name = string("op_2049_transpose_x_0"), val = bool(false)];
            bool var_2049_transpose_y_0 = const()[name = string("op_2049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1904_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2049_cast_fp16 = matmul(transpose_x = var_2049_transpose_x_0, transpose_y = var_2049_transpose_y_0, x = var_2048_cast_fp16, y = v_3_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2049_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2066_begin_0 = const()[name = string("op_2066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2066_end_0 = const()[name = string("op_2066_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2066_end_mask_0 = const()[name = string("op_2066_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2066_cast_fp16 = slice_by_index(begin = var_2066_begin_0, end = var_2066_end_0, end_mask = var_2066_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<int32, [4]> var_2081_begin_0 = const()[name = string("op_2081_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2081_end_0 = const()[name = string("op_2081_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2081_end_mask_0 = const()[name = string("op_2081_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2081_cast_fp16 = slice_by_index(begin = var_2081_begin_0, end = var_2081_end_0, end_mask = var_2081_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2081_cast_fp16")];
            int32 var_2083 = const()[name = string("op_2083"), val = int32(1)];
            bool mixed_7_interleave_0 = const()[name = string("mixed_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_7_cast_fp16 = concat(axis = var_2083, interleave = mixed_7_interleave_0, values = (var_2066_cast_fp16, var_2081_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2087_perm_0 = const()[name = string("op_2087_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2091 = const()[name = string("op_2091"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2087_cast_fp16 = transpose(perm = var_2087_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2091, x = var_2087_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331213248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333310464))))[name = string("model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = add(x = x_13_cast_fp16, y = linear_10_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_2102_promoted_to_fp16 = const()[name = string("op_2102_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2104_cast_fp16 = pow(x = x_23_cast_fp16, y = var_2102_promoted_to_fp16)[name = string("op_2104_cast_fp16")];
            tensor<int32, [1]> var_2106_axes_0 = const()[name = string("op_2106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2106_keep_dims_0 = const()[name = string("op_2106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2106_cast_fp16 = reduce_mean(axes = var_2106_axes_0, keep_dims = var_2106_keep_dims_0, x = var_2104_cast_fp16)[name = string("op_2106_cast_fp16")];
            fp16 var_2107_to_fp16 = const()[name = string("op_2107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2108_cast_fp16 = add(x = var_2106_cast_fp16, y = var_2107_to_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2110_cast_fp16 = mul(x = x_23_cast_fp16, y = var_2109_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333312576)))];
            tensor<fp16, [1, 16, 1024]> input_15_cast_fp16 = mul(x = var_2110_cast_fp16, y = model_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333314688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336460480))))[name = string("model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2115_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336466688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339612480))))[name = string("model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_19_cast_fp16 = mul(x = var_2115_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339618688))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342764480))))[name = string("model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_13_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_2129_promoted_to_fp16 = const()[name = string("op_2129_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2131_cast_fp16 = pow(x = x_25_cast_fp16, y = var_2129_promoted_to_fp16)[name = string("op_2131_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2133_keep_dims_0 = const()[name = string("op_2133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2133_cast_fp16 = reduce_mean(axes = var_2133_axes_0, keep_dims = var_2133_keep_dims_0, x = var_2131_cast_fp16)[name = string("op_2133_cast_fp16")];
            fp16 var_2134_to_fp16 = const()[name = string("op_2134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2135_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = string("op_2135_cast_fp16")];
            fp32 var_2136_epsilon_0 = const()[name = string("op_2136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2136_cast_fp16 = rsqrt(epsilon = var_2136_epsilon_0, x = var_2135_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2137_cast_fp16 = mul(x = x_25_cast_fp16, y = var_2136_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342766592)))];
            tensor<fp16, [1, 16, 1024]> input_21_cast_fp16 = mul(x = var_2137_cast_fp16, y = model_layers_2_input_layernorm_weight_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342768704))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344865920))))[name = string("model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_2146 = const()[name = string("op_2146"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2147_cast_fp16 = reshape(shape = var_2146, x = linear_14_cast_fp16)[name = string("op_2147_cast_fp16")];
            tensor<int32, [4]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2155_promoted_to_fp16 = const()[name = string("op_2155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = var_2147_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2157_cast_fp16 = pow(x = x_27_cast_fp16, y = var_2155_promoted_to_fp16)[name = string("op_2157_cast_fp16")];
            tensor<int32, [1]> var_2159_axes_0 = const()[name = string("op_2159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2159_keep_dims_0 = const()[name = string("op_2159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2159_cast_fp16 = reduce_mean(axes = var_2159_axes_0, keep_dims = var_2159_keep_dims_0, x = var_2157_cast_fp16)[name = string("op_2159_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2161_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
            fp32 var_2162_epsilon_0 = const()[name = string("op_2162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2162_cast_fp16 = rsqrt(epsilon = var_2162_epsilon_0, x = var_2161_cast_fp16)[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2163_cast_fp16 = mul(x = x_27_cast_fp16, y = var_2162_cast_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870080)))];
            tensor<fp16, [1, 16, 16, 128]> x_31_cast_fp16 = mul(x = var_2163_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345919040))))[name = string("model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2172 = const()[name = string("op_2172"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2173_cast_fp16 = reshape(shape = var_2172, x = linear_15_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2181_promoted_to_fp16 = const()[name = string("op_2181_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_2173_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2183_cast_fp16 = pow(x = x_29_cast_fp16, y = var_2181_promoted_to_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [1]> var_2185_axes_0 = const()[name = string("op_2185_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2185_keep_dims_0 = const()[name = string("op_2185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2185_cast_fp16 = reduce_mean(axes = var_2185_axes_0, keep_dims = var_2185_keep_dims_0, x = var_2183_cast_fp16)[name = string("op_2185_cast_fp16")];
            fp16 var_2186_to_fp16 = const()[name = string("op_2186_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2187_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2186_to_fp16)[name = string("op_2187_cast_fp16")];
            fp32 var_2188_epsilon_0 = const()[name = string("op_2188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2188_cast_fp16 = rsqrt(epsilon = var_2188_epsilon_0, x = var_2187_cast_fp16)[name = string("op_2188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2189_cast_fp16 = mul(x = x_29_cast_fp16, y = var_2188_cast_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921152)))];
            tensor<fp16, [1, 8, 16, 128]> x_33_cast_fp16 = mul(x = var_2189_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346970112))))[name = string("model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2198 = const()[name = string("op_2198"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2199_cast_fp16 = reshape(shape = var_2198, x = linear_16_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_9_begin_0 = const()[name = string("first_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_9_end_0 = const()[name = string("first_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_9_end_mask_0 = const()[name = string("first_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_9_cast_fp16 = slice_by_index(begin = first_9_begin_0, end = first_9_end_0, end_mask = first_9_end_mask_0, x = x_31_cast_fp16)[name = string("first_9_cast_fp16")];
            tensor<int32, [4]> second_9_begin_0 = const()[name = string("second_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_9_end_0 = const()[name = string("second_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_9_end_mask_0 = const()[name = string("second_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_9_cast_fp16 = slice_by_index(begin = second_9_begin_0, end = second_9_end_0, end_mask = second_9_end_mask_0, x = x_31_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2213_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2213_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2214_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2216_cast_fp16 = sub(x = var_2213_cast_fp16, y = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2217_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2218_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2220_cast_fp16 = add(x = var_2217_cast_fp16, y = var_2218_cast_fp16)[name = string("op_2220_cast_fp16")];
            int32 var_2222 = const()[name = string("op_2222"), val = int32(-1)];
            bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_5_cast_fp16 = concat(axis = var_2222, interleave = q_5_interleave_0, values = (var_2216_cast_fp16, var_2220_cast_fp16))[name = string("q_5_cast_fp16")];
            tensor<int32, [4]> first_11_begin_0 = const()[name = string("first_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_11_end_0 = const()[name = string("first_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_11_end_mask_0 = const()[name = string("first_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_11_cast_fp16 = slice_by_index(begin = first_11_begin_0, end = first_11_end_0, end_mask = first_11_end_mask_0, x = x_33_cast_fp16)[name = string("first_11_cast_fp16")];
            tensor<int32, [4]> second_11_begin_0 = const()[name = string("second_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_11_end_0 = const()[name = string("second_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_11_end_mask_0 = const()[name = string("second_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_11_cast_fp16 = slice_by_index(begin = second_11_begin_0, end = second_11_end_0, end_mask = second_11_end_mask_0, x = x_33_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2234_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2235_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2237_cast_fp16 = sub(x = var_2234_cast_fp16, y = var_2235_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2238_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2238_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2239_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2241_cast_fp16 = add(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = string("op_2241_cast_fp16")];
            int32 var_2243 = const()[name = string("op_2243"), val = int32(-1)];
            bool k_5_interleave_0 = const()[name = string("k_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_5_cast_fp16 = concat(axis = var_2243, interleave = k_5_interleave_0, values = (var_2237_cast_fp16, var_2241_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2254_begin_0 = const()[name = string("op_2254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2254_end_0 = const()[name = string("op_2254_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2254_end_mask_0 = const()[name = string("op_2254_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2254_stride_0 = const()[name = string("op_2254_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2254_cast_fp16 = slice_by_index(begin = var_2254_begin_0, end = var_2254_end_0, end_mask = var_2254_end_mask_0, stride = var_2254_stride_0, x = q_5_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<int32, [4]> var_2264_begin_0 = const()[name = string("op_2264_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2264_end_0 = const()[name = string("op_2264_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2264_end_mask_0 = const()[name = string("op_2264_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2264_stride_0 = const()[name = string("op_2264_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2264_cast_fp16 = slice_by_index(begin = var_2264_begin_0, end = var_2264_end_0, end_mask = var_2264_end_mask_0, stride = var_2264_stride_0, x = q_5_cast_fp16)[name = string("op_2264_cast_fp16")];
            int32 var_2266 = const()[name = string("op_2266"), val = int32(2)];
            bool stacked_5_interleave_0 = const()[name = string("stacked_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_5_cast_fp16 = concat(axis = var_2266, interleave = stacked_5_interleave_0, values = (var_2254_cast_fp16, var_2264_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2272_begin_0 = const()[name = string("op_2272_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2272_end_0 = const()[name = string("op_2272_end_0"), val = tensor<int32, [4]>([3, 8, 2064, 128])];
            tensor<bool, [4]> var_2272_end_mask_0 = const()[name = string("op_2272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2272_cast_fp16 = slice_by_index(begin = var_2272_begin_0, end = var_2272_end_0, end_mask = var_2272_end_mask_0, x = read_state_0)[name = string("op_2272_cast_fp16")];
            tensor<int32, [4]> var_2282_begin_0 = const()[name = string("op_2282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2282_end_0 = const()[name = string("op_2282_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2282_end_mask_0 = const()[name = string("op_2282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2282_cast_fp16 = slice_by_index(begin = var_2282_begin_0, end = var_2282_end_0, end_mask = var_2282_end_mask_0, x = var_2272_cast_fp16)[name = string("op_2282_cast_fp16")];
            bool var_2291_transpose_x_1 = const()[name = string("op_2291_transpose_x_1"), val = bool(false)];
            bool var_2291_transpose_y_1 = const()[name = string("op_2291_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2291_cast_fp16 = matmul(transpose_x = var_2291_transpose_x_1, transpose_y = var_2291_transpose_y_1, x = var_2282_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2294_perm_0 = const()[name = string("op_2294_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2295_to_fp16 = const()[name = string("op_2295_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2294_cast_fp16 = transpose(perm = var_2294_perm_0, x = var_2291_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 2048]> var_2296_cast_fp16 = mul(x = var_2294_cast_fp16, y = var_2295_to_fp16)[name = string("op_2296_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_5_cast_fp16 = add(x = var_2296_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2302_transpose_x_1 = const()[name = string("op_2302_transpose_x_1"), val = bool(false)];
            bool var_2302_transpose_y_1 = const()[name = string("op_2302_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2302_cast_fp16 = matmul(transpose_x = var_2302_transpose_x_1, transpose_y = var_2302_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp16 var_2303_to_fp16 = const()[name = string("op_2303_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2304_cast_fp16 = mul(x = var_2302_cast_fp16, y = var_2303_to_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2304_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2308 = const()[name = string("op_2308"), val = int32(-1)];
            bool var_2309_interleave_0 = const()[name = string("op_2309_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2309_cast_fp16 = concat(axis = var_2308, interleave = var_2309_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2309_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_5_cast_fp16 = softmax(axis = var_2310, x = var_2309_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<int32, [4]> var_2322_begin_0 = const()[name = string("op_2322_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2322_end_0 = const()[name = string("op_2322_end_0"), val = tensor<int32, [4]>([3, 8, 2064, 128])];
            tensor<bool, [4]> var_2322_end_mask_0 = const()[name = string("op_2322_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = read_state_1)[name = string("op_2322_cast_fp16")];
            tensor<int32, [4]> var_2332_begin_0 = const()[name = string("op_2332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2332_end_0 = const()[name = string("op_2332_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2332_end_mask_0 = const()[name = string("op_2332_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2332_cast_fp16 = slice_by_index(begin = var_2332_begin_0, end = var_2332_end_0, end_mask = var_2332_end_mask_0, x = var_2322_cast_fp16)[name = string("op_2332_cast_fp16")];
            bool var_2338_transpose_x_0 = const()[name = string("op_2338_transpose_x_0"), val = bool(false)];
            bool var_2338_transpose_y_0 = const()[name = string("op_2338_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2338_cast_fp16 = matmul(transpose_x = var_2338_transpose_x_0, transpose_y = var_2338_transpose_y_0, x = var_2317_cast_fp16, y = var_2332_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<int32, [4]> var_2343_begin_0 = const()[name = string("op_2343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2343_end_0 = const()[name = string("op_2343_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2343_end_mask_0 = const()[name = string("op_2343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2343_cast_fp16 = slice_by_index(begin = var_2343_begin_0, end = var_2343_end_0, end_mask = var_2343_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2343_cast_fp16")];
            bool var_2344_transpose_x_0 = const()[name = string("op_2344_transpose_x_0"), val = bool(false)];
            bool var_2344_transpose_y_0 = const()[name = string("op_2344_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2199_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2344_cast_fp16 = matmul(transpose_x = var_2344_transpose_x_0, transpose_y = var_2344_transpose_y_0, x = var_2343_cast_fp16, y = v_5_cast_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2338_cast_fp16, y = var_2344_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2361_begin_0 = const()[name = string("op_2361_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2361_end_0 = const()[name = string("op_2361_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2361_end_mask_0 = const()[name = string("op_2361_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2361_cast_fp16")];
            tensor<int32, [4]> var_2376_begin_0 = const()[name = string("op_2376_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2376_end_0 = const()[name = string("op_2376_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2376_end_mask_0 = const()[name = string("op_2376_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2376_cast_fp16 = slice_by_index(begin = var_2376_begin_0, end = var_2376_end_0, end_mask = var_2376_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(1)];
            bool mixed_11_interleave_0 = const()[name = string("mixed_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_11_cast_fp16 = concat(axis = var_2378, interleave = mixed_11_interleave_0, values = (var_2361_cast_fp16, var_2376_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2382_perm_0 = const()[name = string("op_2382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2382_cast_fp16 = transpose(perm = var_2382_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346972224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349069440))))[name = string("model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_35_cast_fp16 = add(x = x_25_cast_fp16, y = linear_17_cast_fp16)[name = string("x_35_cast_fp16")];
            fp16 var_2397_promoted_to_fp16 = const()[name = string("op_2397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2399_cast_fp16 = pow(x = x_35_cast_fp16, y = var_2397_promoted_to_fp16)[name = string("op_2399_cast_fp16")];
            tensor<int32, [1]> var_2401_axes_0 = const()[name = string("op_2401_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2401_keep_dims_0 = const()[name = string("op_2401_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2401_cast_fp16 = reduce_mean(axes = var_2401_axes_0, keep_dims = var_2401_keep_dims_0, x = var_2399_cast_fp16)[name = string("op_2401_cast_fp16")];
            fp16 var_2402_to_fp16 = const()[name = string("op_2402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2403_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2402_to_fp16)[name = string("op_2403_cast_fp16")];
            fp32 var_2404_epsilon_0 = const()[name = string("op_2404_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2404_cast_fp16 = rsqrt(epsilon = var_2404_epsilon_0, x = var_2403_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2405_cast_fp16 = mul(x = x_35_cast_fp16, y = var_2404_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349071552)))];
            tensor<fp16, [1, 16, 1024]> input_25_cast_fp16 = mul(x = var_2405_cast_fp16, y = model_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349073664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352219456))))[name = string("model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2410_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352225664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355371456))))[name = string("model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_29_cast_fp16 = mul(x = var_2410_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355377664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358523456))))[name = string("model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_20_cast_fp16)[name = string("x_37_cast_fp16")];
            fp16 var_2424_promoted_to_fp16 = const()[name = string("op_2424_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2426_cast_fp16 = pow(x = x_37_cast_fp16, y = var_2424_promoted_to_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [1]> var_2428_axes_0 = const()[name = string("op_2428_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2428_keep_dims_0 = const()[name = string("op_2428_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2428_cast_fp16 = reduce_mean(axes = var_2428_axes_0, keep_dims = var_2428_keep_dims_0, x = var_2426_cast_fp16)[name = string("op_2428_cast_fp16")];
            fp16 var_2429_to_fp16 = const()[name = string("op_2429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2430_cast_fp16 = add(x = var_2428_cast_fp16, y = var_2429_to_fp16)[name = string("op_2430_cast_fp16")];
            fp32 var_2431_epsilon_0 = const()[name = string("op_2431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2431_cast_fp16 = rsqrt(epsilon = var_2431_epsilon_0, x = var_2430_cast_fp16)[name = string("op_2431_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2432_cast_fp16 = mul(x = x_37_cast_fp16, y = var_2431_cast_fp16)[name = string("op_2432_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358525568)))];
            tensor<fp16, [1, 16, 1024]> input_31_cast_fp16 = mul(x = var_2432_cast_fp16, y = model_layers_3_input_layernorm_weight_to_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358527680))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360624896))))[name = string("model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2442_cast_fp16 = reshape(shape = var_2441, x = linear_21_cast_fp16)[name = string("op_2442_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2450_promoted_to_fp16 = const()[name = string("op_2450_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2442_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2452_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2450_promoted_to_fp16)[name = string("op_2452_cast_fp16")];
            tensor<int32, [1]> var_2454_axes_0 = const()[name = string("op_2454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2454_keep_dims_0 = const()[name = string("op_2454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2454_cast_fp16 = reduce_mean(axes = var_2454_axes_0, keep_dims = var_2454_keep_dims_0, x = var_2452_cast_fp16)[name = string("op_2454_cast_fp16")];
            fp16 var_2455_to_fp16 = const()[name = string("op_2455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2456_cast_fp16 = add(x = var_2454_cast_fp16, y = var_2455_to_fp16)[name = string("op_2456_cast_fp16")];
            fp32 var_2457_epsilon_0 = const()[name = string("op_2457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2457_cast_fp16 = rsqrt(epsilon = var_2457_epsilon_0, x = var_2456_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2458_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629056)))];
            tensor<fp16, [1, 16, 16, 128]> x_43_cast_fp16 = mul(x = var_2458_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361678016))))[name = string("model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2468_cast_fp16 = reshape(shape = var_2467, x = linear_22_cast_fp16)[name = string("op_2468_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2476_promoted_to_fp16 = const()[name = string("op_2476_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2468_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2478_cast_fp16 = pow(x = x_41_cast_fp16, y = var_2476_promoted_to_fp16)[name = string("op_2478_cast_fp16")];
            tensor<int32, [1]> var_2480_axes_0 = const()[name = string("op_2480_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2480_keep_dims_0 = const()[name = string("op_2480_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2480_cast_fp16 = reduce_mean(axes = var_2480_axes_0, keep_dims = var_2480_keep_dims_0, x = var_2478_cast_fp16)[name = string("op_2480_cast_fp16")];
            fp16 var_2481_to_fp16 = const()[name = string("op_2481_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2482_cast_fp16 = add(x = var_2480_cast_fp16, y = var_2481_to_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2484_cast_fp16 = mul(x = x_41_cast_fp16, y = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680128)))];
            tensor<fp16, [1, 8, 16, 128]> x_45_cast_fp16 = mul(x = var_2484_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362729088))))[name = string("model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2493 = const()[name = string("op_2493"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2494_cast_fp16 = reshape(shape = var_2493, x = linear_23_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_13_begin_0 = const()[name = string("first_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_13_end_0 = const()[name = string("first_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_13_end_mask_0 = const()[name = string("first_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_13_cast_fp16 = slice_by_index(begin = first_13_begin_0, end = first_13_end_0, end_mask = first_13_end_mask_0, x = x_43_cast_fp16)[name = string("first_13_cast_fp16")];
            tensor<int32, [4]> second_13_begin_0 = const()[name = string("second_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_13_end_0 = const()[name = string("second_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_13_end_mask_0 = const()[name = string("second_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_13_cast_fp16 = slice_by_index(begin = second_13_begin_0, end = second_13_end_0, end_mask = second_13_end_mask_0, x = x_43_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2508_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2508_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2509_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2509_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2511_cast_fp16 = sub(x = var_2508_cast_fp16, y = var_2509_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2512_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2512_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2513_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2515_cast_fp16 = add(x = var_2512_cast_fp16, y = var_2513_cast_fp16)[name = string("op_2515_cast_fp16")];
            int32 var_2517 = const()[name = string("op_2517"), val = int32(-1)];
            bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_7_cast_fp16 = concat(axis = var_2517, interleave = q_7_interleave_0, values = (var_2511_cast_fp16, var_2515_cast_fp16))[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> first_15_begin_0 = const()[name = string("first_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_15_end_0 = const()[name = string("first_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_15_end_mask_0 = const()[name = string("first_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_15_cast_fp16 = slice_by_index(begin = first_15_begin_0, end = first_15_end_0, end_mask = first_15_end_mask_0, x = x_45_cast_fp16)[name = string("first_15_cast_fp16")];
            tensor<int32, [4]> second_15_begin_0 = const()[name = string("second_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_15_end_0 = const()[name = string("second_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_15_end_mask_0 = const()[name = string("second_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_15_cast_fp16 = slice_by_index(begin = second_15_begin_0, end = second_15_end_0, end_mask = second_15_end_mask_0, x = x_45_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2529_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2530_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2532_cast_fp16 = sub(x = var_2529_cast_fp16, y = var_2530_cast_fp16)[name = string("op_2532_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2533_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2534_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2534_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2536_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2536_cast_fp16")];
            int32 var_2538 = const()[name = string("op_2538"), val = int32(-1)];
            bool k_7_interleave_0 = const()[name = string("k_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_7_cast_fp16 = concat(axis = var_2538, interleave = k_7_interleave_0, values = (var_2532_cast_fp16, var_2536_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2549_begin_0 = const()[name = string("op_2549_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2549_end_0 = const()[name = string("op_2549_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2549_end_mask_0 = const()[name = string("op_2549_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2549_stride_0 = const()[name = string("op_2549_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2549_cast_fp16 = slice_by_index(begin = var_2549_begin_0, end = var_2549_end_0, end_mask = var_2549_end_mask_0, stride = var_2549_stride_0, x = q_7_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [4]> var_2559_begin_0 = const()[name = string("op_2559_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2559_end_0 = const()[name = string("op_2559_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2559_end_mask_0 = const()[name = string("op_2559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2559_stride_0 = const()[name = string("op_2559_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2559_cast_fp16 = slice_by_index(begin = var_2559_begin_0, end = var_2559_end_0, end_mask = var_2559_end_mask_0, stride = var_2559_stride_0, x = q_7_cast_fp16)[name = string("op_2559_cast_fp16")];
            int32 var_2561 = const()[name = string("op_2561"), val = int32(2)];
            bool stacked_7_interleave_0 = const()[name = string("stacked_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_7_cast_fp16 = concat(axis = var_2561, interleave = stacked_7_interleave_0, values = (var_2549_cast_fp16, var_2559_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2567_begin_0 = const()[name = string("op_2567_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2567_end_0 = const()[name = string("op_2567_end_0"), val = tensor<int32, [4]>([4, 8, 2064, 128])];
            tensor<bool, [4]> var_2567_end_mask_0 = const()[name = string("op_2567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2567_cast_fp16 = slice_by_index(begin = var_2567_begin_0, end = var_2567_end_0, end_mask = var_2567_end_mask_0, x = read_state_0)[name = string("op_2567_cast_fp16")];
            tensor<int32, [4]> var_2577_begin_0 = const()[name = string("op_2577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2577_end_0 = const()[name = string("op_2577_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2577_end_mask_0 = const()[name = string("op_2577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2577_cast_fp16 = slice_by_index(begin = var_2577_begin_0, end = var_2577_end_0, end_mask = var_2577_end_mask_0, x = var_2567_cast_fp16)[name = string("op_2577_cast_fp16")];
            bool var_2586_transpose_x_1 = const()[name = string("op_2586_transpose_x_1"), val = bool(false)];
            bool var_2586_transpose_y_1 = const()[name = string("op_2586_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2586_cast_fp16 = matmul(transpose_x = var_2586_transpose_x_1, transpose_y = var_2586_transpose_y_1, x = var_2577_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2586_cast_fp16")];
            tensor<int32, [4]> var_2589_perm_0 = const()[name = string("op_2589_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2590_to_fp16 = const()[name = string("op_2590_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2589_cast_fp16 = transpose(perm = var_2589_perm_0, x = var_2586_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 2048]> var_2591_cast_fp16 = mul(x = var_2589_cast_fp16, y = var_2590_to_fp16)[name = string("op_2591_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_7_cast_fp16 = add(x = var_2591_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2597_transpose_x_1 = const()[name = string("op_2597_transpose_x_1"), val = bool(false)];
            bool var_2597_transpose_y_1 = const()[name = string("op_2597_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2597_cast_fp16 = matmul(transpose_x = var_2597_transpose_x_1, transpose_y = var_2597_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2597_cast_fp16")];
            fp16 var_2598_to_fp16 = const()[name = string("op_2598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2599_cast_fp16 = mul(x = var_2597_cast_fp16, y = var_2598_to_fp16)[name = string("op_2599_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2599_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(-1)];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2604_cast_fp16 = concat(axis = var_2603, interleave = var_2604_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2604_cast_fp16")];
            int32 var_2605 = const()[name = string("op_2605"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_7_cast_fp16 = softmax(axis = var_2605, x = var_2604_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2612_begin_0 = const()[name = string("op_2612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2612_end_0 = const()[name = string("op_2612_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2612_end_mask_0 = const()[name = string("op_2612_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2612_cast_fp16 = slice_by_index(begin = var_2612_begin_0, end = var_2612_end_0, end_mask = var_2612_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2612_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([4, 8, 2064, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = read_state_1)[name = string("op_2617_cast_fp16")];
            tensor<int32, [4]> var_2627_begin_0 = const()[name = string("op_2627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2627_end_0 = const()[name = string("op_2627_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2627_end_mask_0 = const()[name = string("op_2627_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2627_cast_fp16 = slice_by_index(begin = var_2627_begin_0, end = var_2627_end_0, end_mask = var_2627_end_mask_0, x = var_2617_cast_fp16)[name = string("op_2627_cast_fp16")];
            bool var_2633_transpose_x_0 = const()[name = string("op_2633_transpose_x_0"), val = bool(false)];
            bool var_2633_transpose_y_0 = const()[name = string("op_2633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2633_cast_fp16 = matmul(transpose_x = var_2633_transpose_x_0, transpose_y = var_2633_transpose_y_0, x = var_2612_cast_fp16, y = var_2627_cast_fp16)[name = string("op_2633_cast_fp16")];
            tensor<int32, [4]> var_2638_begin_0 = const()[name = string("op_2638_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2638_end_0 = const()[name = string("op_2638_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2638_end_mask_0 = const()[name = string("op_2638_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2638_cast_fp16 = slice_by_index(begin = var_2638_begin_0, end = var_2638_end_0, end_mask = var_2638_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2638_cast_fp16")];
            bool var_2639_transpose_x_0 = const()[name = string("op_2639_transpose_x_0"), val = bool(false)];
            bool var_2639_transpose_y_0 = const()[name = string("op_2639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2494_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2639_cast_fp16 = matmul(transpose_x = var_2639_transpose_x_0, transpose_y = var_2639_transpose_y_0, x = var_2638_cast_fp16, y = v_7_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2633_cast_fp16, y = var_2639_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2656_cast_fp16")];
            tensor<int32, [4]> var_2671_begin_0 = const()[name = string("op_2671_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2671_end_0 = const()[name = string("op_2671_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2671_end_mask_0 = const()[name = string("op_2671_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2671_cast_fp16 = slice_by_index(begin = var_2671_begin_0, end = var_2671_end_0, end_mask = var_2671_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2671_cast_fp16")];
            int32 var_2673 = const()[name = string("op_2673"), val = int32(1)];
            bool mixed_15_interleave_0 = const()[name = string("mixed_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_15_cast_fp16 = concat(axis = var_2673, interleave = mixed_15_interleave_0, values = (var_2656_cast_fp16, var_2671_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2681 = const()[name = string("op_2681"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2681, x = var_2677_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362731200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364828416))))[name = string("model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = linear_24_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2692_promoted_to_fp16 = const()[name = string("op_2692_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2694_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2692_promoted_to_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [1]> var_2696_axes_0 = const()[name = string("op_2696_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2696_keep_dims_0 = const()[name = string("op_2696_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2696_cast_fp16 = reduce_mean(axes = var_2696_axes_0, keep_dims = var_2696_keep_dims_0, x = var_2694_cast_fp16)[name = string("op_2696_cast_fp16")];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2698_cast_fp16 = add(x = var_2696_cast_fp16, y = var_2697_to_fp16)[name = string("op_2698_cast_fp16")];
            fp32 var_2699_epsilon_0 = const()[name = string("op_2699_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2699_cast_fp16 = rsqrt(epsilon = var_2699_epsilon_0, x = var_2698_cast_fp16)[name = string("op_2699_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2700_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2699_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364830528)))];
            tensor<fp16, [1, 16, 1024]> input_35_cast_fp16 = mul(x = var_2700_cast_fp16, y = model_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364832640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367978432))))[name = string("model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2705_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367984640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371130432))))[name = string("model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_39_cast_fp16 = mul(x = var_2705_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371136640))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374282432))))[name = string("model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_27_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_2719_promoted_to_fp16 = const()[name = string("op_2719_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2721_cast_fp16 = pow(x = x_49_cast_fp16, y = var_2719_promoted_to_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]>([-1])];
            bool var_2723_keep_dims_0 = const()[name = string("op_2723_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2723_cast_fp16 = reduce_mean(axes = var_2723_axes_0, keep_dims = var_2723_keep_dims_0, x = var_2721_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 var_2724_to_fp16 = const()[name = string("op_2724_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2725_cast_fp16 = add(x = var_2723_cast_fp16, y = var_2724_to_fp16)[name = string("op_2725_cast_fp16")];
            fp32 var_2726_epsilon_0 = const()[name = string("op_2726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2726_cast_fp16 = rsqrt(epsilon = var_2726_epsilon_0, x = var_2725_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2727_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2726_cast_fp16)[name = string("op_2727_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374284544)))];
            tensor<fp16, [1, 16, 1024]> input_41_cast_fp16 = mul(x = var_2727_cast_fp16, y = model_layers_4_input_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374286656))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376383872))))[name = string("model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_2736 = const()[name = string("op_2736"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2737_cast_fp16 = reshape(shape = var_2736, x = linear_28_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<int32, [4]> x_51_perm_0 = const()[name = string("x_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2745_promoted_to_fp16 = const()[name = string("op_2745_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_51_cast_fp16 = transpose(perm = x_51_perm_0, x = var_2737_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2747_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2745_promoted_to_fp16)[name = string("op_2747_cast_fp16")];
            tensor<int32, [1]> var_2749_axes_0 = const()[name = string("op_2749_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2749_keep_dims_0 = const()[name = string("op_2749_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2749_cast_fp16 = reduce_mean(axes = var_2749_axes_0, keep_dims = var_2749_keep_dims_0, x = var_2747_cast_fp16)[name = string("op_2749_cast_fp16")];
            fp16 var_2750_to_fp16 = const()[name = string("op_2750_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2751_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2750_to_fp16)[name = string("op_2751_cast_fp16")];
            fp32 var_2752_epsilon_0 = const()[name = string("op_2752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2752_cast_fp16 = rsqrt(epsilon = var_2752_epsilon_0, x = var_2751_cast_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2753_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2752_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388032)))];
            tensor<fp16, [1, 16, 16, 128]> x_55_cast_fp16 = mul(x = var_2753_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377436992))))[name = string("model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2762 = const()[name = string("op_2762"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2763_cast_fp16 = reshape(shape = var_2762, x = linear_29_cast_fp16)[name = string("op_2763_cast_fp16")];
            tensor<int32, [4]> x_53_perm_0 = const()[name = string("x_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2771_promoted_to_fp16 = const()[name = string("op_2771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_53_cast_fp16 = transpose(perm = x_53_perm_0, x = var_2763_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2773_cast_fp16 = pow(x = x_53_cast_fp16, y = var_2771_promoted_to_fp16)[name = string("op_2773_cast_fp16")];
            tensor<int32, [1]> var_2775_axes_0 = const()[name = string("op_2775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2775_keep_dims_0 = const()[name = string("op_2775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2775_cast_fp16 = reduce_mean(axes = var_2775_axes_0, keep_dims = var_2775_keep_dims_0, x = var_2773_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp16 var_2776_to_fp16 = const()[name = string("op_2776_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2777_cast_fp16 = add(x = var_2775_cast_fp16, y = var_2776_to_fp16)[name = string("op_2777_cast_fp16")];
            fp32 var_2778_epsilon_0 = const()[name = string("op_2778_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2778_cast_fp16 = rsqrt(epsilon = var_2778_epsilon_0, x = var_2777_cast_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2779_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2779_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439104)))];
            tensor<fp16, [1, 8, 16, 128]> x_57_cast_fp16 = mul(x = var_2779_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488064))))[name = string("model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2788 = const()[name = string("op_2788"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2789_cast_fp16 = reshape(shape = var_2788, x = linear_30_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_17_begin_0 = const()[name = string("first_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_17_end_0 = const()[name = string("first_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_17_end_mask_0 = const()[name = string("first_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_17_cast_fp16 = slice_by_index(begin = first_17_begin_0, end = first_17_end_0, end_mask = first_17_end_mask_0, x = x_55_cast_fp16)[name = string("first_17_cast_fp16")];
            tensor<int32, [4]> second_17_begin_0 = const()[name = string("second_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_17_end_0 = const()[name = string("second_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_17_end_mask_0 = const()[name = string("second_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_17_cast_fp16 = slice_by_index(begin = second_17_begin_0, end = second_17_end_0, end_mask = second_17_end_mask_0, x = x_55_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2803_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2804_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2804_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2806_cast_fp16 = sub(x = var_2803_cast_fp16, y = var_2804_cast_fp16)[name = string("op_2806_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2807_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2808_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2808_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2810_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2808_cast_fp16)[name = string("op_2810_cast_fp16")];
            int32 var_2812 = const()[name = string("op_2812"), val = int32(-1)];
            bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_9_cast_fp16 = concat(axis = var_2812, interleave = q_9_interleave_0, values = (var_2806_cast_fp16, var_2810_cast_fp16))[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> first_19_begin_0 = const()[name = string("first_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_19_end_0 = const()[name = string("first_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_19_end_mask_0 = const()[name = string("first_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_19_cast_fp16 = slice_by_index(begin = first_19_begin_0, end = first_19_end_0, end_mask = first_19_end_mask_0, x = x_57_cast_fp16)[name = string("first_19_cast_fp16")];
            tensor<int32, [4]> second_19_begin_0 = const()[name = string("second_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_19_end_0 = const()[name = string("second_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_19_end_mask_0 = const()[name = string("second_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_19_cast_fp16 = slice_by_index(begin = second_19_begin_0, end = second_19_end_0, end_mask = second_19_end_mask_0, x = x_57_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2824_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2824_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2825_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2827_cast_fp16 = sub(x = var_2824_cast_fp16, y = var_2825_cast_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2828_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2829_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2831_cast_fp16 = add(x = var_2828_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2831_cast_fp16")];
            int32 var_2833 = const()[name = string("op_2833"), val = int32(-1)];
            bool k_9_interleave_0 = const()[name = string("k_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_9_cast_fp16 = concat(axis = var_2833, interleave = k_9_interleave_0, values = (var_2827_cast_fp16, var_2831_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2844_begin_0 = const()[name = string("op_2844_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2844_end_0 = const()[name = string("op_2844_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2844_end_mask_0 = const()[name = string("op_2844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2844_stride_0 = const()[name = string("op_2844_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, stride = var_2844_stride_0, x = q_9_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<int32, [4]> var_2854_begin_0 = const()[name = string("op_2854_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2854_end_0 = const()[name = string("op_2854_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2854_end_mask_0 = const()[name = string("op_2854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2854_stride_0 = const()[name = string("op_2854_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2854_cast_fp16 = slice_by_index(begin = var_2854_begin_0, end = var_2854_end_0, end_mask = var_2854_end_mask_0, stride = var_2854_stride_0, x = q_9_cast_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(2)];
            bool stacked_9_interleave_0 = const()[name = string("stacked_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_9_cast_fp16 = concat(axis = var_2856, interleave = stacked_9_interleave_0, values = (var_2844_cast_fp16, var_2854_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2862_begin_0 = const()[name = string("op_2862_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2862_end_0 = const()[name = string("op_2862_end_0"), val = tensor<int32, [4]>([5, 8, 2064, 128])];
            tensor<bool, [4]> var_2862_end_mask_0 = const()[name = string("op_2862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2862_cast_fp16 = slice_by_index(begin = var_2862_begin_0, end = var_2862_end_0, end_mask = var_2862_end_mask_0, x = read_state_0)[name = string("op_2862_cast_fp16")];
            tensor<int32, [4]> var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = var_2862_cast_fp16)[name = string("op_2872_cast_fp16")];
            bool var_2881_transpose_x_1 = const()[name = string("op_2881_transpose_x_1"), val = bool(false)];
            bool var_2881_transpose_y_1 = const()[name = string("op_2881_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_2881_cast_fp16 = matmul(transpose_x = var_2881_transpose_x_1, transpose_y = var_2881_transpose_y_1, x = var_2872_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_2881_cast_fp16")];
            tensor<int32, [4]> var_2884_perm_0 = const()[name = string("op_2884_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2885_to_fp16 = const()[name = string("op_2885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_2884_cast_fp16 = transpose(perm = var_2884_perm_0, x = var_2881_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 2048]> var_2886_cast_fp16 = mul(x = var_2884_cast_fp16, y = var_2885_to_fp16)[name = string("op_2886_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_9_cast_fp16 = add(x = var_2886_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_2892_transpose_x_1 = const()[name = string("op_2892_transpose_x_1"), val = bool(false)];
            bool var_2892_transpose_y_1 = const()[name = string("op_2892_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2892_cast_fp16 = matmul(transpose_x = var_2892_transpose_x_1, transpose_y = var_2892_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_2892_cast_fp16")];
            fp16 var_2893_to_fp16 = const()[name = string("op_2893_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2894_cast_fp16 = mul(x = var_2892_cast_fp16, y = var_2893_to_fp16)[name = string("op_2894_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_2894_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_2898 = const()[name = string("op_2898"), val = int32(-1)];
            bool var_2899_interleave_0 = const()[name = string("op_2899_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_2899_cast_fp16 = concat(axis = var_2898, interleave = var_2899_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_2899_cast_fp16")];
            int32 var_2900 = const()[name = string("op_2900"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_9_cast_fp16 = softmax(axis = var_2900, x = var_2899_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_2907_begin_0 = const()[name = string("op_2907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2907_end_0 = const()[name = string("op_2907_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_2907_end_mask_0 = const()[name = string("op_2907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_2907_cast_fp16 = slice_by_index(begin = var_2907_begin_0, end = var_2907_end_0, end_mask = var_2907_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<int32, [4]> var_2912_begin_0 = const()[name = string("op_2912_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2912_end_0 = const()[name = string("op_2912_end_0"), val = tensor<int32, [4]>([5, 8, 2064, 128])];
            tensor<bool, [4]> var_2912_end_mask_0 = const()[name = string("op_2912_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_2912_cast_fp16 = slice_by_index(begin = var_2912_begin_0, end = var_2912_end_0, end_mask = var_2912_end_mask_0, x = read_state_1)[name = string("op_2912_cast_fp16")];
            tensor<int32, [4]> var_2922_begin_0 = const()[name = string("op_2922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2922_end_0 = const()[name = string("op_2922_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_2922_end_mask_0 = const()[name = string("op_2922_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_2922_cast_fp16 = slice_by_index(begin = var_2922_begin_0, end = var_2922_end_0, end_mask = var_2922_end_mask_0, x = var_2912_cast_fp16)[name = string("op_2922_cast_fp16")];
            bool var_2928_transpose_x_0 = const()[name = string("op_2928_transpose_x_0"), val = bool(false)];
            bool var_2928_transpose_y_0 = const()[name = string("op_2928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2928_cast_fp16 = matmul(transpose_x = var_2928_transpose_x_0, transpose_y = var_2928_transpose_y_0, x = var_2907_cast_fp16, y = var_2922_cast_fp16)[name = string("op_2928_cast_fp16")];
            tensor<int32, [4]> var_2933_begin_0 = const()[name = string("op_2933_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_2933_end_0 = const()[name = string("op_2933_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_2933_end_mask_0 = const()[name = string("op_2933_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2933_cast_fp16 = slice_by_index(begin = var_2933_begin_0, end = var_2933_end_0, end_mask = var_2933_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2933_cast_fp16")];
            bool var_2934_transpose_x_0 = const()[name = string("op_2934_transpose_x_0"), val = bool(false)];
            bool var_2934_transpose_y_0 = const()[name = string("op_2934_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2789_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_2934_cast_fp16 = matmul(transpose_x = var_2934_transpose_x_0, transpose_y = var_2934_transpose_y_0, x = var_2933_cast_fp16, y = v_9_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2934_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_2951_begin_0 = const()[name = string("op_2951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2951_end_0 = const()[name = string("op_2951_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2951_end_mask_0 = const()[name = string("op_2951_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2951_cast_fp16 = slice_by_index(begin = var_2951_begin_0, end = var_2951_end_0, end_mask = var_2951_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2951_cast_fp16")];
            tensor<int32, [4]> var_2966_begin_0 = const()[name = string("op_2966_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2966_end_0 = const()[name = string("op_2966_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2966_end_mask_0 = const()[name = string("op_2966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2966_cast_fp16 = slice_by_index(begin = var_2966_begin_0, end = var_2966_end_0, end_mask = var_2966_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2966_cast_fp16")];
            int32 var_2968 = const()[name = string("op_2968"), val = int32(1)];
            bool mixed_19_interleave_0 = const()[name = string("mixed_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_19_cast_fp16 = concat(axis = var_2968, interleave = mixed_19_interleave_0, values = (var_2951_cast_fp16, var_2966_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_2972_perm_0 = const()[name = string("op_2972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2976 = const()[name = string("op_2976"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2972_cast_fp16 = transpose(perm = var_2972_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_2976, x = var_2972_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378490176))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380587392))))[name = string("model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = linear_31_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2987_promoted_to_fp16 = const()[name = string("op_2987_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2989_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2987_promoted_to_fp16)[name = string("op_2989_cast_fp16")];
            tensor<int32, [1]> var_2991_axes_0 = const()[name = string("op_2991_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2991_keep_dims_0 = const()[name = string("op_2991_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2991_cast_fp16 = reduce_mean(axes = var_2991_axes_0, keep_dims = var_2991_keep_dims_0, x = var_2989_cast_fp16)[name = string("op_2991_cast_fp16")];
            fp16 var_2992_to_fp16 = const()[name = string("op_2992_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2993_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_to_fp16)[name = string("op_2993_cast_fp16")];
            fp32 var_2994_epsilon_0 = const()[name = string("op_2994_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2994_cast_fp16 = rsqrt(epsilon = var_2994_epsilon_0, x = var_2993_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2995_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2994_cast_fp16)[name = string("op_2995_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380589504)))];
            tensor<fp16, [1, 16, 1024]> input_45_cast_fp16 = mul(x = var_2995_cast_fp16, y = model_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380591616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383737408))))[name = string("model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3000_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3000_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383743616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889408))))[name = string("model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_49_cast_fp16 = mul(x = var_3000_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386895616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390041408))))[name = string("model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_34_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_3014_promoted_to_fp16 = const()[name = string("op_3014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3016_cast_fp16 = pow(x = x_61_cast_fp16, y = var_3014_promoted_to_fp16)[name = string("op_3016_cast_fp16")];
            tensor<int32, [1]> var_3018_axes_0 = const()[name = string("op_3018_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3018_keep_dims_0 = const()[name = string("op_3018_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3018_cast_fp16 = reduce_mean(axes = var_3018_axes_0, keep_dims = var_3018_keep_dims_0, x = var_3016_cast_fp16)[name = string("op_3018_cast_fp16")];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3020_cast_fp16 = add(x = var_3018_cast_fp16, y = var_3019_to_fp16)[name = string("op_3020_cast_fp16")];
            fp32 var_3021_epsilon_0 = const()[name = string("op_3021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3021_cast_fp16 = rsqrt(epsilon = var_3021_epsilon_0, x = var_3020_cast_fp16)[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3022_cast_fp16 = mul(x = x_61_cast_fp16, y = var_3021_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390043520)))];
            tensor<fp16, [1, 16, 1024]> input_51_cast_fp16 = mul(x = var_3022_cast_fp16, y = model_layers_5_input_layernorm_weight_to_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390045632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392142848))))[name = string("model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3032_cast_fp16 = reshape(shape = var_3031, x = linear_35_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> x_63_perm_0 = const()[name = string("x_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3040_promoted_to_fp16 = const()[name = string("op_3040_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_63_cast_fp16 = transpose(perm = x_63_perm_0, x = var_3032_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3042_cast_fp16 = pow(x = x_63_cast_fp16, y = var_3040_promoted_to_fp16)[name = string("op_3042_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3042_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 var_3045_to_fp16 = const()[name = string("op_3045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3046_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3045_to_fp16)[name = string("op_3046_cast_fp16")];
            fp32 var_3047_epsilon_0 = const()[name = string("op_3047_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3047_cast_fp16 = rsqrt(epsilon = var_3047_epsilon_0, x = var_3046_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3048_cast_fp16 = mul(x = x_63_cast_fp16, y = var_3047_cast_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147008)))];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = mul(x = var_3048_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393195968))))[name = string("model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3058_cast_fp16 = reshape(shape = var_3057, x = linear_36_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3066_promoted_to_fp16 = const()[name = string("op_3066_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_3058_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3068_cast_fp16 = pow(x = x_65_cast_fp16, y = var_3066_promoted_to_fp16)[name = string("op_3068_cast_fp16")];
            tensor<int32, [1]> var_3070_axes_0 = const()[name = string("op_3070_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3070_keep_dims_0 = const()[name = string("op_3070_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3070_cast_fp16 = reduce_mean(axes = var_3070_axes_0, keep_dims = var_3070_keep_dims_0, x = var_3068_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp16 var_3071_to_fp16 = const()[name = string("op_3071_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3072_cast_fp16 = add(x = var_3070_cast_fp16, y = var_3071_to_fp16)[name = string("op_3072_cast_fp16")];
            fp32 var_3073_epsilon_0 = const()[name = string("op_3073_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3073_cast_fp16 = rsqrt(epsilon = var_3073_epsilon_0, x = var_3072_cast_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3074_cast_fp16 = mul(x = x_65_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198080)))];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = mul(x = var_3074_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394247040))))[name = string("model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3083 = const()[name = string("op_3083"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3084_cast_fp16 = reshape(shape = var_3083, x = linear_37_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_21_begin_0 = const()[name = string("first_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_21_end_0 = const()[name = string("first_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_21_end_mask_0 = const()[name = string("first_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_21_cast_fp16 = slice_by_index(begin = first_21_begin_0, end = first_21_end_0, end_mask = first_21_end_mask_0, x = x_67_cast_fp16)[name = string("first_21_cast_fp16")];
            tensor<int32, [4]> second_21_begin_0 = const()[name = string("second_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_21_end_0 = const()[name = string("second_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_21_end_mask_0 = const()[name = string("second_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_21_cast_fp16 = slice_by_index(begin = second_21_begin_0, end = second_21_end_0, end_mask = second_21_end_mask_0, x = x_67_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3098_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3098_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3099_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3101_cast_fp16 = sub(x = var_3098_cast_fp16, y = var_3099_cast_fp16)[name = string("op_3101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3102_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3102_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3103_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3105_cast_fp16 = add(x = var_3102_cast_fp16, y = var_3103_cast_fp16)[name = string("op_3105_cast_fp16")];
            int32 var_3107 = const()[name = string("op_3107"), val = int32(-1)];
            bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_11_cast_fp16 = concat(axis = var_3107, interleave = q_11_interleave_0, values = (var_3101_cast_fp16, var_3105_cast_fp16))[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> first_23_begin_0 = const()[name = string("first_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_23_end_0 = const()[name = string("first_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_23_end_mask_0 = const()[name = string("first_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_23_cast_fp16 = slice_by_index(begin = first_23_begin_0, end = first_23_end_0, end_mask = first_23_end_mask_0, x = x_69_cast_fp16)[name = string("first_23_cast_fp16")];
            tensor<int32, [4]> second_23_begin_0 = const()[name = string("second_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_23_end_0 = const()[name = string("second_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_23_end_mask_0 = const()[name = string("second_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_23_cast_fp16 = slice_by_index(begin = second_23_begin_0, end = second_23_end_0, end_mask = second_23_end_mask_0, x = x_69_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3119_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3119_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3120_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3122_cast_fp16 = sub(x = var_3119_cast_fp16, y = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3123_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3123_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3124_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3124_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3126_cast_fp16 = add(x = var_3123_cast_fp16, y = var_3124_cast_fp16)[name = string("op_3126_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-1)];
            bool k_11_interleave_0 = const()[name = string("k_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_11_cast_fp16 = concat(axis = var_3128, interleave = k_11_interleave_0, values = (var_3122_cast_fp16, var_3126_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3139_begin_0 = const()[name = string("op_3139_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3139_end_0 = const()[name = string("op_3139_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3139_end_mask_0 = const()[name = string("op_3139_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3139_stride_0 = const()[name = string("op_3139_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3139_cast_fp16 = slice_by_index(begin = var_3139_begin_0, end = var_3139_end_0, end_mask = var_3139_end_mask_0, stride = var_3139_stride_0, x = q_11_cast_fp16)[name = string("op_3139_cast_fp16")];
            tensor<int32, [4]> var_3149_begin_0 = const()[name = string("op_3149_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3149_end_0 = const()[name = string("op_3149_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3149_end_mask_0 = const()[name = string("op_3149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3149_stride_0 = const()[name = string("op_3149_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3149_cast_fp16 = slice_by_index(begin = var_3149_begin_0, end = var_3149_end_0, end_mask = var_3149_end_mask_0, stride = var_3149_stride_0, x = q_11_cast_fp16)[name = string("op_3149_cast_fp16")];
            int32 var_3151 = const()[name = string("op_3151"), val = int32(2)];
            bool stacked_11_interleave_0 = const()[name = string("stacked_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_11_cast_fp16 = concat(axis = var_3151, interleave = stacked_11_interleave_0, values = (var_3139_cast_fp16, var_3149_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3157_begin_0 = const()[name = string("op_3157_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3157_end_0 = const()[name = string("op_3157_end_0"), val = tensor<int32, [4]>([6, 8, 2064, 128])];
            tensor<bool, [4]> var_3157_end_mask_0 = const()[name = string("op_3157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3157_cast_fp16 = slice_by_index(begin = var_3157_begin_0, end = var_3157_end_0, end_mask = var_3157_end_mask_0, x = read_state_0)[name = string("op_3157_cast_fp16")];
            tensor<int32, [4]> var_3167_begin_0 = const()[name = string("op_3167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3167_end_0 = const()[name = string("op_3167_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3167_end_mask_0 = const()[name = string("op_3167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, x = var_3157_cast_fp16)[name = string("op_3167_cast_fp16")];
            bool var_3176_transpose_x_1 = const()[name = string("op_3176_transpose_x_1"), val = bool(false)];
            bool var_3176_transpose_y_1 = const()[name = string("op_3176_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3176_cast_fp16 = matmul(transpose_x = var_3176_transpose_x_1, transpose_y = var_3176_transpose_y_1, x = var_3167_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3176_cast_fp16")];
            tensor<int32, [4]> var_3179_perm_0 = const()[name = string("op_3179_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3180_to_fp16 = const()[name = string("op_3180_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3179_cast_fp16 = transpose(perm = var_3179_perm_0, x = var_3176_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 2048]> var_3181_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_3180_to_fp16)[name = string("op_3181_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_11_cast_fp16 = add(x = var_3181_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3187_transpose_x_1 = const()[name = string("op_3187_transpose_x_1"), val = bool(false)];
            bool var_3187_transpose_y_1 = const()[name = string("op_3187_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3187_cast_fp16 = matmul(transpose_x = var_3187_transpose_x_1, transpose_y = var_3187_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3187_cast_fp16")];
            fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3189_cast_fp16 = mul(x = var_3187_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3189_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3193 = const()[name = string("op_3193"), val = int32(-1)];
            bool var_3194_interleave_0 = const()[name = string("op_3194_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3194_cast_fp16 = concat(axis = var_3193, interleave = var_3194_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3194_cast_fp16")];
            int32 var_3195 = const()[name = string("op_3195"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_11_cast_fp16 = softmax(axis = var_3195, x = var_3194_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3202_begin_0 = const()[name = string("op_3202_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3202_end_0 = const()[name = string("op_3202_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3202_end_mask_0 = const()[name = string("op_3202_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3202_cast_fp16 = slice_by_index(begin = var_3202_begin_0, end = var_3202_end_0, end_mask = var_3202_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3202_cast_fp16")];
            tensor<int32, [4]> var_3207_begin_0 = const()[name = string("op_3207_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3207_end_0 = const()[name = string("op_3207_end_0"), val = tensor<int32, [4]>([6, 8, 2064, 128])];
            tensor<bool, [4]> var_3207_end_mask_0 = const()[name = string("op_3207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3207_cast_fp16 = slice_by_index(begin = var_3207_begin_0, end = var_3207_end_0, end_mask = var_3207_end_mask_0, x = read_state_1)[name = string("op_3207_cast_fp16")];
            tensor<int32, [4]> var_3217_begin_0 = const()[name = string("op_3217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3217_end_0 = const()[name = string("op_3217_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3217_end_mask_0 = const()[name = string("op_3217_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3217_cast_fp16 = slice_by_index(begin = var_3217_begin_0, end = var_3217_end_0, end_mask = var_3217_end_mask_0, x = var_3207_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3223_transpose_x_0 = const()[name = string("op_3223_transpose_x_0"), val = bool(false)];
            bool var_3223_transpose_y_0 = const()[name = string("op_3223_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3223_cast_fp16 = matmul(transpose_x = var_3223_transpose_x_0, transpose_y = var_3223_transpose_y_0, x = var_3202_cast_fp16, y = var_3217_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<int32, [4]> var_3228_begin_0 = const()[name = string("op_3228_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3228_end_0 = const()[name = string("op_3228_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3228_end_mask_0 = const()[name = string("op_3228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3228_cast_fp16 = slice_by_index(begin = var_3228_begin_0, end = var_3228_end_0, end_mask = var_3228_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3229_transpose_x_0 = const()[name = string("op_3229_transpose_x_0"), val = bool(false)];
            bool var_3229_transpose_y_0 = const()[name = string("op_3229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3084_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3229_cast_fp16 = matmul(transpose_x = var_3229_transpose_x_0, transpose_y = var_3229_transpose_y_0, x = var_3228_cast_fp16, y = v_11_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3246_begin_0 = const()[name = string("op_3246_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3246_end_0 = const()[name = string("op_3246_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3246_end_mask_0 = const()[name = string("op_3246_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3246_cast_fp16 = slice_by_index(begin = var_3246_begin_0, end = var_3246_end_0, end_mask = var_3246_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3246_cast_fp16")];
            tensor<int32, [4]> var_3261_begin_0 = const()[name = string("op_3261_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3261_end_0 = const()[name = string("op_3261_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3261_end_mask_0 = const()[name = string("op_3261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3261_cast_fp16 = slice_by_index(begin = var_3261_begin_0, end = var_3261_end_0, end_mask = var_3261_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3261_cast_fp16")];
            int32 var_3263 = const()[name = string("op_3263"), val = int32(1)];
            bool mixed_23_interleave_0 = const()[name = string("mixed_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_23_cast_fp16 = concat(axis = var_3263, interleave = mixed_23_interleave_0, values = (var_3246_cast_fp16, var_3261_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3267_perm_0 = const()[name = string("op_3267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3267_cast_fp16 = transpose(perm = var_3267_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3271, x = var_3267_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394249152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396346368))))[name = string("model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_71_cast_fp16 = add(x = x_61_cast_fp16, y = linear_38_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_3282_promoted_to_fp16 = const()[name = string("op_3282_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3284_cast_fp16 = pow(x = x_71_cast_fp16, y = var_3282_promoted_to_fp16)[name = string("op_3284_cast_fp16")];
            tensor<int32, [1]> var_3286_axes_0 = const()[name = string("op_3286_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3286_keep_dims_0 = const()[name = string("op_3286_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3286_cast_fp16 = reduce_mean(axes = var_3286_axes_0, keep_dims = var_3286_keep_dims_0, x = var_3284_cast_fp16)[name = string("op_3286_cast_fp16")];
            fp16 var_3287_to_fp16 = const()[name = string("op_3287_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3288_cast_fp16 = add(x = var_3286_cast_fp16, y = var_3287_to_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3289_cast_fp16 = rsqrt(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3290_cast_fp16 = mul(x = x_71_cast_fp16, y = var_3289_cast_fp16)[name = string("op_3290_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396348480)))];
            tensor<fp16, [1, 16, 1024]> input_55_cast_fp16 = mul(x = var_3290_cast_fp16, y = model_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396350592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399496384))))[name = string("model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3295_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3295_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399502592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402648384))))[name = string("model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_59_cast_fp16 = mul(x = var_3295_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402654592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405800384))))[name = string("model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_41_cast_fp16)[name = string("x_73_cast_fp16")];
            fp16 var_3309_promoted_to_fp16 = const()[name = string("op_3309_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3311_cast_fp16 = pow(x = x_73_cast_fp16, y = var_3309_promoted_to_fp16)[name = string("op_3311_cast_fp16")];
            tensor<int32, [1]> var_3313_axes_0 = const()[name = string("op_3313_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3313_keep_dims_0 = const()[name = string("op_3313_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3313_cast_fp16 = reduce_mean(axes = var_3313_axes_0, keep_dims = var_3313_keep_dims_0, x = var_3311_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3315_cast_fp16 = add(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            fp32 var_3316_epsilon_0 = const()[name = string("op_3316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3316_cast_fp16 = rsqrt(epsilon = var_3316_epsilon_0, x = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3317_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405802496)))];
            tensor<fp16, [1, 16, 1024]> input_61_cast_fp16 = mul(x = var_3317_cast_fp16, y = model_layers_6_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405804608))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407901824))))[name = string("model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_3326 = const()[name = string("op_3326"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3327_cast_fp16 = reshape(shape = var_3326, x = linear_42_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3335_promoted_to_fp16 = const()[name = string("op_3335_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3327_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3337_cast_fp16 = pow(x = x_75_cast_fp16, y = var_3335_promoted_to_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]>([-1])];
            bool var_3339_keep_dims_0 = const()[name = string("op_3339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3339_cast_fp16 = reduce_mean(axes = var_3339_axes_0, keep_dims = var_3339_keep_dims_0, x = var_3337_cast_fp16)[name = string("op_3339_cast_fp16")];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3341_cast_fp16 = add(x = var_3339_cast_fp16, y = var_3340_to_fp16)[name = string("op_3341_cast_fp16")];
            fp32 var_3342_epsilon_0 = const()[name = string("op_3342_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3342_cast_fp16 = rsqrt(epsilon = var_3342_epsilon_0, x = var_3341_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3343_cast_fp16 = mul(x = x_75_cast_fp16, y = var_3342_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407905984)))];
            tensor<fp16, [1, 16, 16, 128]> x_79_cast_fp16 = mul(x = var_3343_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407906304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408954944))))[name = string("model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3352 = const()[name = string("op_3352"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3353_cast_fp16 = reshape(shape = var_3352, x = linear_43_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3361_promoted_to_fp16 = const()[name = string("op_3361_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3353_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3363_cast_fp16 = pow(x = x_77_cast_fp16, y = var_3361_promoted_to_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3365_keep_dims_0 = const()[name = string("op_3365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3365_cast_fp16 = reduce_mean(axes = var_3365_axes_0, keep_dims = var_3365_keep_dims_0, x = var_3363_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3367_cast_fp16 = add(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("op_3367_cast_fp16")];
            fp32 var_3368_epsilon_0 = const()[name = string("op_3368_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3368_cast_fp16 = rsqrt(epsilon = var_3368_epsilon_0, x = var_3367_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3369_cast_fp16 = mul(x = x_77_cast_fp16, y = var_3368_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957056)))];
            tensor<fp16, [1, 8, 16, 128]> x_81_cast_fp16 = mul(x = var_3369_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410006016))))[name = string("model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3378 = const()[name = string("op_3378"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3379_cast_fp16 = reshape(shape = var_3378, x = linear_44_cast_fp16)[name = string("op_3379_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_25_begin_0 = const()[name = string("first_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_25_end_0 = const()[name = string("first_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_25_end_mask_0 = const()[name = string("first_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_25_cast_fp16 = slice_by_index(begin = first_25_begin_0, end = first_25_end_0, end_mask = first_25_end_mask_0, x = x_79_cast_fp16)[name = string("first_25_cast_fp16")];
            tensor<int32, [4]> second_25_begin_0 = const()[name = string("second_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_25_end_0 = const()[name = string("second_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_25_end_mask_0 = const()[name = string("second_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_25_cast_fp16 = slice_by_index(begin = second_25_begin_0, end = second_25_end_0, end_mask = second_25_end_mask_0, x = x_79_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3393_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3393_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3394_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3396_cast_fp16 = sub(x = var_3393_cast_fp16, y = var_3394_cast_fp16)[name = string("op_3396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3397_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3398_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_13_cast_fp16 = concat(axis = var_3402, interleave = q_13_interleave_0, values = (var_3396_cast_fp16, var_3400_cast_fp16))[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> first_27_begin_0 = const()[name = string("first_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_27_end_0 = const()[name = string("first_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_27_end_mask_0 = const()[name = string("first_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_27_cast_fp16 = slice_by_index(begin = first_27_begin_0, end = first_27_end_0, end_mask = first_27_end_mask_0, x = x_81_cast_fp16)[name = string("first_27_cast_fp16")];
            tensor<int32, [4]> second_27_begin_0 = const()[name = string("second_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_27_end_0 = const()[name = string("second_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_27_end_mask_0 = const()[name = string("second_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_27_cast_fp16 = slice_by_index(begin = second_27_begin_0, end = second_27_end_0, end_mask = second_27_end_mask_0, x = x_81_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3414_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3414_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3415_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3415_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3417_cast_fp16 = sub(x = var_3414_cast_fp16, y = var_3415_cast_fp16)[name = string("op_3417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3418_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3419_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3421_cast_fp16 = add(x = var_3418_cast_fp16, y = var_3419_cast_fp16)[name = string("op_3421_cast_fp16")];
            int32 var_3423 = const()[name = string("op_3423"), val = int32(-1)];
            bool k_13_interleave_0 = const()[name = string("k_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_13_cast_fp16 = concat(axis = var_3423, interleave = k_13_interleave_0, values = (var_3417_cast_fp16, var_3421_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3434_begin_0 = const()[name = string("op_3434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3434_end_0 = const()[name = string("op_3434_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3434_end_mask_0 = const()[name = string("op_3434_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3434_stride_0 = const()[name = string("op_3434_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3434_cast_fp16 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, stride = var_3434_stride_0, x = q_13_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<int32, [4]> var_3444_begin_0 = const()[name = string("op_3444_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3444_end_0 = const()[name = string("op_3444_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3444_end_mask_0 = const()[name = string("op_3444_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3444_stride_0 = const()[name = string("op_3444_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3444_cast_fp16 = slice_by_index(begin = var_3444_begin_0, end = var_3444_end_0, end_mask = var_3444_end_mask_0, stride = var_3444_stride_0, x = q_13_cast_fp16)[name = string("op_3444_cast_fp16")];
            int32 var_3446 = const()[name = string("op_3446"), val = int32(2)];
            bool stacked_13_interleave_0 = const()[name = string("stacked_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_13_cast_fp16 = concat(axis = var_3446, interleave = stacked_13_interleave_0, values = (var_3434_cast_fp16, var_3444_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3452_begin_0 = const()[name = string("op_3452_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3452_end_0 = const()[name = string("op_3452_end_0"), val = tensor<int32, [4]>([7, 8, 2064, 128])];
            tensor<bool, [4]> var_3452_end_mask_0 = const()[name = string("op_3452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3452_cast_fp16 = slice_by_index(begin = var_3452_begin_0, end = var_3452_end_0, end_mask = var_3452_end_mask_0, x = read_state_0)[name = string("op_3452_cast_fp16")];
            tensor<int32, [4]> var_3462_begin_0 = const()[name = string("op_3462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3462_end_0 = const()[name = string("op_3462_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3462_end_mask_0 = const()[name = string("op_3462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3462_cast_fp16 = slice_by_index(begin = var_3462_begin_0, end = var_3462_end_0, end_mask = var_3462_end_mask_0, x = var_3452_cast_fp16)[name = string("op_3462_cast_fp16")];
            bool var_3471_transpose_x_1 = const()[name = string("op_3471_transpose_x_1"), val = bool(false)];
            bool var_3471_transpose_y_1 = const()[name = string("op_3471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3471_cast_fp16 = matmul(transpose_x = var_3471_transpose_x_1, transpose_y = var_3471_transpose_y_1, x = var_3462_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<int32, [4]> var_3474_perm_0 = const()[name = string("op_3474_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3475_to_fp16 = const()[name = string("op_3475_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3474_cast_fp16 = transpose(perm = var_3474_perm_0, x = var_3471_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 2048]> var_3476_cast_fp16 = mul(x = var_3474_cast_fp16, y = var_3475_to_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_13_cast_fp16 = add(x = var_3476_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3482_transpose_x_1 = const()[name = string("op_3482_transpose_x_1"), val = bool(false)];
            bool var_3482_transpose_y_1 = const()[name = string("op_3482_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3482_cast_fp16 = matmul(transpose_x = var_3482_transpose_x_1, transpose_y = var_3482_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3482_cast_fp16")];
            fp16 var_3483_to_fp16 = const()[name = string("op_3483_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3484_cast_fp16 = mul(x = var_3482_cast_fp16, y = var_3483_to_fp16)[name = string("op_3484_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3484_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3488 = const()[name = string("op_3488"), val = int32(-1)];
            bool var_3489_interleave_0 = const()[name = string("op_3489_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3489_cast_fp16 = concat(axis = var_3488, interleave = var_3489_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3489_cast_fp16")];
            int32 var_3490 = const()[name = string("op_3490"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_13_cast_fp16 = softmax(axis = var_3490, x = var_3489_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3497_begin_0 = const()[name = string("op_3497_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3497_end_0 = const()[name = string("op_3497_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3497_end_mask_0 = const()[name = string("op_3497_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3497_cast_fp16 = slice_by_index(begin = var_3497_begin_0, end = var_3497_end_0, end_mask = var_3497_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> var_3502_begin_0 = const()[name = string("op_3502_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3502_end_0 = const()[name = string("op_3502_end_0"), val = tensor<int32, [4]>([7, 8, 2064, 128])];
            tensor<bool, [4]> var_3502_end_mask_0 = const()[name = string("op_3502_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3502_cast_fp16 = slice_by_index(begin = var_3502_begin_0, end = var_3502_end_0, end_mask = var_3502_end_mask_0, x = read_state_1)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> var_3512_begin_0 = const()[name = string("op_3512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3512_end_0 = const()[name = string("op_3512_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3512_end_mask_0 = const()[name = string("op_3512_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3512_cast_fp16 = slice_by_index(begin = var_3512_begin_0, end = var_3512_end_0, end_mask = var_3512_end_mask_0, x = var_3502_cast_fp16)[name = string("op_3512_cast_fp16")];
            bool var_3518_transpose_x_0 = const()[name = string("op_3518_transpose_x_0"), val = bool(false)];
            bool var_3518_transpose_y_0 = const()[name = string("op_3518_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3518_cast_fp16 = matmul(transpose_x = var_3518_transpose_x_0, transpose_y = var_3518_transpose_y_0, x = var_3497_cast_fp16, y = var_3512_cast_fp16)[name = string("op_3518_cast_fp16")];
            tensor<int32, [4]> var_3523_begin_0 = const()[name = string("op_3523_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3523_end_0 = const()[name = string("op_3523_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3523_end_mask_0 = const()[name = string("op_3523_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3523_cast_fp16 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3523_cast_fp16")];
            bool var_3524_transpose_x_0 = const()[name = string("op_3524_transpose_x_0"), val = bool(false)];
            bool var_3524_transpose_y_0 = const()[name = string("op_3524_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3379_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3524_cast_fp16 = matmul(transpose_x = var_3524_transpose_x_0, transpose_y = var_3524_transpose_y_0, x = var_3523_cast_fp16, y = v_13_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3524_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3541_begin_0 = const()[name = string("op_3541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3541_end_0 = const()[name = string("op_3541_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3541_end_mask_0 = const()[name = string("op_3541_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3541_cast_fp16 = slice_by_index(begin = var_3541_begin_0, end = var_3541_end_0, end_mask = var_3541_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3541_cast_fp16")];
            tensor<int32, [4]> var_3556_begin_0 = const()[name = string("op_3556_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3556_end_0 = const()[name = string("op_3556_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3556_end_mask_0 = const()[name = string("op_3556_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3556_cast_fp16 = slice_by_index(begin = var_3556_begin_0, end = var_3556_end_0, end_mask = var_3556_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3556_cast_fp16")];
            int32 var_3558 = const()[name = string("op_3558"), val = int32(1)];
            bool mixed_27_interleave_0 = const()[name = string("mixed_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_27_cast_fp16 = concat(axis = var_3558, interleave = mixed_27_interleave_0, values = (var_3541_cast_fp16, var_3556_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3562_perm_0 = const()[name = string("op_3562_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3562_cast_fp16 = transpose(perm = var_3562_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3566, x = var_3562_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410008128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412105344))))[name = string("model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = linear_45_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_3577_promoted_to_fp16 = const()[name = string("op_3577_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3579_cast_fp16 = pow(x = x_83_cast_fp16, y = var_3577_promoted_to_fp16)[name = string("op_3579_cast_fp16")];
            tensor<int32, [1]> var_3581_axes_0 = const()[name = string("op_3581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3581_keep_dims_0 = const()[name = string("op_3581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3581_cast_fp16 = reduce_mean(axes = var_3581_axes_0, keep_dims = var_3581_keep_dims_0, x = var_3579_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp16 var_3582_to_fp16 = const()[name = string("op_3582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3583_cast_fp16 = add(x = var_3581_cast_fp16, y = var_3582_to_fp16)[name = string("op_3583_cast_fp16")];
            fp32 var_3584_epsilon_0 = const()[name = string("op_3584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3584_cast_fp16 = rsqrt(epsilon = var_3584_epsilon_0, x = var_3583_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3585_cast_fp16 = mul(x = x_83_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412107456)))];
            tensor<fp16, [1, 16, 1024]> input_65_cast_fp16 = mul(x = var_3585_cast_fp16, y = model_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412109568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415255360))))[name = string("model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3590_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3590_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415261568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418407360))))[name = string("model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_69_cast_fp16 = mul(x = var_3590_cast_fp16, y = linear_47_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418413568))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421559360))))[name = string("model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_48_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_3604_promoted_to_fp16 = const()[name = string("op_3604_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3606_cast_fp16 = pow(x = x_85_cast_fp16, y = var_3604_promoted_to_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3608_keep_dims_0 = const()[name = string("op_3608_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3608_cast_fp16 = reduce_mean(axes = var_3608_axes_0, keep_dims = var_3608_keep_dims_0, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3610_cast_fp16 = add(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("op_3610_cast_fp16")];
            fp32 var_3611_epsilon_0 = const()[name = string("op_3611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3611_cast_fp16 = rsqrt(epsilon = var_3611_epsilon_0, x = var_3610_cast_fp16)[name = string("op_3611_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3612_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3611_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421561472)))];
            tensor<fp16, [1, 16, 1024]> input_71_cast_fp16 = mul(x = var_3612_cast_fp16, y = model_layers_7_input_layernorm_weight_to_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421563584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423660800))))[name = string("model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3622_cast_fp16 = reshape(shape = var_3621, x = linear_49_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3630_promoted_to_fp16 = const()[name = string("op_3630_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_3622_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3632_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3630_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [1]> var_3634_axes_0 = const()[name = string("op_3634_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3634_keep_dims_0 = const()[name = string("op_3634_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3634_cast_fp16 = reduce_mean(axes = var_3634_axes_0, keep_dims = var_3634_keep_dims_0, x = var_3632_cast_fp16)[name = string("op_3634_cast_fp16")];
            fp16 var_3635_to_fp16 = const()[name = string("op_3635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3636_cast_fp16 = add(x = var_3634_cast_fp16, y = var_3635_to_fp16)[name = string("op_3636_cast_fp16")];
            fp32 var_3637_epsilon_0 = const()[name = string("op_3637_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3637_cast_fp16 = rsqrt(epsilon = var_3637_epsilon_0, x = var_3636_cast_fp16)[name = string("op_3637_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3638_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3637_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423664960)))];
            tensor<fp16, [1, 16, 16, 128]> x_91_cast_fp16 = mul(x = var_3638_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423665280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424713920))))[name = string("model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3647 = const()[name = string("op_3647"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3648_cast_fp16 = reshape(shape = var_3647, x = linear_50_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3656_promoted_to_fp16 = const()[name = string("op_3656_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_3648_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3658_cast_fp16 = pow(x = x_89_cast_fp16, y = var_3656_promoted_to_fp16)[name = string("op_3658_cast_fp16")];
            tensor<int32, [1]> var_3660_axes_0 = const()[name = string("op_3660_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3660_keep_dims_0 = const()[name = string("op_3660_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3660_cast_fp16 = reduce_mean(axes = var_3660_axes_0, keep_dims = var_3660_keep_dims_0, x = var_3658_cast_fp16)[name = string("op_3660_cast_fp16")];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3662_cast_fp16 = add(x = var_3660_cast_fp16, y = var_3661_to_fp16)[name = string("op_3662_cast_fp16")];
            fp32 var_3663_epsilon_0 = const()[name = string("op_3663_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3663_cast_fp16 = rsqrt(epsilon = var_3663_epsilon_0, x = var_3662_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3664_cast_fp16 = mul(x = x_89_cast_fp16, y = var_3663_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716032)))];
            tensor<fp16, [1, 8, 16, 128]> x_93_cast_fp16 = mul(x = var_3664_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425764992))))[name = string("model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3673 = const()[name = string("op_3673"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3674_cast_fp16 = reshape(shape = var_3673, x = linear_51_cast_fp16)[name = string("op_3674_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_29_begin_0 = const()[name = string("first_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_29_end_0 = const()[name = string("first_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_29_end_mask_0 = const()[name = string("first_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_29_cast_fp16 = slice_by_index(begin = first_29_begin_0, end = first_29_end_0, end_mask = first_29_end_mask_0, x = x_91_cast_fp16)[name = string("first_29_cast_fp16")];
            tensor<int32, [4]> second_29_begin_0 = const()[name = string("second_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_29_end_0 = const()[name = string("second_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_29_end_mask_0 = const()[name = string("second_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_29_cast_fp16 = slice_by_index(begin = second_29_begin_0, end = second_29_end_0, end_mask = second_29_end_mask_0, x = x_91_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3688_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3689_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3691_cast_fp16 = sub(x = var_3688_cast_fp16, y = var_3689_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3692_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3693_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3695_cast_fp16 = add(x = var_3692_cast_fp16, y = var_3693_cast_fp16)[name = string("op_3695_cast_fp16")];
            int32 var_3697 = const()[name = string("op_3697"), val = int32(-1)];
            bool q_15_interleave_0 = const()[name = string("q_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_15_cast_fp16 = concat(axis = var_3697, interleave = q_15_interleave_0, values = (var_3691_cast_fp16, var_3695_cast_fp16))[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> first_31_begin_0 = const()[name = string("first_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_31_end_0 = const()[name = string("first_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_31_end_mask_0 = const()[name = string("first_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_31_cast_fp16 = slice_by_index(begin = first_31_begin_0, end = first_31_end_0, end_mask = first_31_end_mask_0, x = x_93_cast_fp16)[name = string("first_31_cast_fp16")];
            tensor<int32, [4]> second_31_begin_0 = const()[name = string("second_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_31_end_0 = const()[name = string("second_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_31_end_mask_0 = const()[name = string("second_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_31_cast_fp16 = slice_by_index(begin = second_31_begin_0, end = second_31_end_0, end_mask = second_31_end_mask_0, x = x_93_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3709_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3709_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3710_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3710_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3712_cast_fp16 = sub(x = var_3709_cast_fp16, y = var_3710_cast_fp16)[name = string("op_3712_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3713_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3714_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3714_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3716_cast_fp16 = add(x = var_3713_cast_fp16, y = var_3714_cast_fp16)[name = string("op_3716_cast_fp16")];
            int32 var_3718 = const()[name = string("op_3718"), val = int32(-1)];
            bool k_15_interleave_0 = const()[name = string("k_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_15_cast_fp16 = concat(axis = var_3718, interleave = k_15_interleave_0, values = (var_3712_cast_fp16, var_3716_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3729_begin_0 = const()[name = string("op_3729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3729_end_0 = const()[name = string("op_3729_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3729_end_mask_0 = const()[name = string("op_3729_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3729_stride_0 = const()[name = string("op_3729_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3729_cast_fp16 = slice_by_index(begin = var_3729_begin_0, end = var_3729_end_0, end_mask = var_3729_end_mask_0, stride = var_3729_stride_0, x = q_15_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [4]> var_3739_begin_0 = const()[name = string("op_3739_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3739_end_0 = const()[name = string("op_3739_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3739_end_mask_0 = const()[name = string("op_3739_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3739_stride_0 = const()[name = string("op_3739_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3739_cast_fp16 = slice_by_index(begin = var_3739_begin_0, end = var_3739_end_0, end_mask = var_3739_end_mask_0, stride = var_3739_stride_0, x = q_15_cast_fp16)[name = string("op_3739_cast_fp16")];
            int32 var_3741 = const()[name = string("op_3741"), val = int32(2)];
            bool stacked_15_interleave_0 = const()[name = string("stacked_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_15_cast_fp16 = concat(axis = var_3741, interleave = stacked_15_interleave_0, values = (var_3729_cast_fp16, var_3739_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3747_begin_0 = const()[name = string("op_3747_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3747_end_0 = const()[name = string("op_3747_end_0"), val = tensor<int32, [4]>([8, 8, 2064, 128])];
            tensor<bool, [4]> var_3747_end_mask_0 = const()[name = string("op_3747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3747_cast_fp16 = slice_by_index(begin = var_3747_begin_0, end = var_3747_end_0, end_mask = var_3747_end_mask_0, x = read_state_0)[name = string("op_3747_cast_fp16")];
            tensor<int32, [4]> var_3757_begin_0 = const()[name = string("op_3757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3757_end_0 = const()[name = string("op_3757_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3757_end_mask_0 = const()[name = string("op_3757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3757_cast_fp16 = slice_by_index(begin = var_3757_begin_0, end = var_3757_end_0, end_mask = var_3757_end_mask_0, x = var_3747_cast_fp16)[name = string("op_3757_cast_fp16")];
            bool var_3766_transpose_x_1 = const()[name = string("op_3766_transpose_x_1"), val = bool(false)];
            bool var_3766_transpose_y_1 = const()[name = string("op_3766_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_3766_cast_fp16 = matmul(transpose_x = var_3766_transpose_x_1, transpose_y = var_3766_transpose_y_1, x = var_3757_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<int32, [4]> var_3769_perm_0 = const()[name = string("op_3769_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3770_to_fp16 = const()[name = string("op_3770_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_3769_cast_fp16 = transpose(perm = var_3769_perm_0, x = var_3766_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 2048]> var_3771_cast_fp16 = mul(x = var_3769_cast_fp16, y = var_3770_to_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_15_cast_fp16 = add(x = var_3771_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3777_transpose_x_1 = const()[name = string("op_3777_transpose_x_1"), val = bool(false)];
            bool var_3777_transpose_y_1 = const()[name = string("op_3777_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3777_cast_fp16 = matmul(transpose_x = var_3777_transpose_x_1, transpose_y = var_3777_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3779_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_3783 = const()[name = string("op_3783"), val = int32(-1)];
            bool var_3784_interleave_0 = const()[name = string("op_3784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_3784_cast_fp16 = concat(axis = var_3783, interleave = var_3784_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_3784_cast_fp16")];
            int32 var_3785 = const()[name = string("op_3785"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_15_cast_fp16 = softmax(axis = var_3785, x = var_3784_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_3792_begin_0 = const()[name = string("op_3792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3792_end_0 = const()[name = string("op_3792_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_3792_end_mask_0 = const()[name = string("op_3792_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<int32, [4]> var_3797_begin_0 = const()[name = string("op_3797_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3797_end_0 = const()[name = string("op_3797_end_0"), val = tensor<int32, [4]>([8, 8, 2064, 128])];
            tensor<bool, [4]> var_3797_end_mask_0 = const()[name = string("op_3797_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_3797_cast_fp16 = slice_by_index(begin = var_3797_begin_0, end = var_3797_end_0, end_mask = var_3797_end_mask_0, x = read_state_1)[name = string("op_3797_cast_fp16")];
            tensor<int32, [4]> var_3807_begin_0 = const()[name = string("op_3807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3807_end_0 = const()[name = string("op_3807_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_3807_end_mask_0 = const()[name = string("op_3807_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_3807_cast_fp16 = slice_by_index(begin = var_3807_begin_0, end = var_3807_end_0, end_mask = var_3807_end_mask_0, x = var_3797_cast_fp16)[name = string("op_3807_cast_fp16")];
            bool var_3813_transpose_x_0 = const()[name = string("op_3813_transpose_x_0"), val = bool(false)];
            bool var_3813_transpose_y_0 = const()[name = string("op_3813_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3813_cast_fp16 = matmul(transpose_x = var_3813_transpose_x_0, transpose_y = var_3813_transpose_y_0, x = var_3792_cast_fp16, y = var_3807_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [4]> var_3818_begin_0 = const()[name = string("op_3818_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_3818_end_0 = const()[name = string("op_3818_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_3818_end_mask_0 = const()[name = string("op_3818_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3818_cast_fp16 = slice_by_index(begin = var_3818_begin_0, end = var_3818_end_0, end_mask = var_3818_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3818_cast_fp16")];
            bool var_3819_transpose_x_0 = const()[name = string("op_3819_transpose_x_0"), val = bool(false)];
            bool var_3819_transpose_y_0 = const()[name = string("op_3819_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3674_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_3819_cast_fp16 = matmul(transpose_x = var_3819_transpose_x_0, transpose_y = var_3819_transpose_y_0, x = var_3818_cast_fp16, y = v_15_cast_fp16)[name = string("op_3819_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_3813_cast_fp16, y = var_3819_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_3836_begin_0 = const()[name = string("op_3836_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3836_end_0 = const()[name = string("op_3836_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3836_end_mask_0 = const()[name = string("op_3836_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<int32, [4]> var_3851_begin_0 = const()[name = string("op_3851_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3851_end_0 = const()[name = string("op_3851_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3851_end_mask_0 = const()[name = string("op_3851_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3851_cast_fp16 = slice_by_index(begin = var_3851_begin_0, end = var_3851_end_0, end_mask = var_3851_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3851_cast_fp16")];
            int32 var_3853 = const()[name = string("op_3853"), val = int32(1)];
            bool mixed_31_interleave_0 = const()[name = string("mixed_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_31_cast_fp16 = concat(axis = var_3853, interleave = mixed_31_interleave_0, values = (var_3836_cast_fp16, var_3851_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_3857_perm_0 = const()[name = string("op_3857_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3857_cast_fp16 = transpose(perm = var_3857_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425767104))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427864320))))[name = string("model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_95_cast_fp16 = add(x = x_85_cast_fp16, y = linear_52_cast_fp16)[name = string("x_95_cast_fp16")];
            fp16 var_3872_promoted_to_fp16 = const()[name = string("op_3872_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3874_cast_fp16 = pow(x = x_95_cast_fp16, y = var_3872_promoted_to_fp16)[name = string("op_3874_cast_fp16")];
            tensor<int32, [1]> var_3876_axes_0 = const()[name = string("op_3876_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3876_keep_dims_0 = const()[name = string("op_3876_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3876_cast_fp16 = reduce_mean(axes = var_3876_axes_0, keep_dims = var_3876_keep_dims_0, x = var_3874_cast_fp16)[name = string("op_3876_cast_fp16")];
            fp16 var_3877_to_fp16 = const()[name = string("op_3877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3878_cast_fp16 = add(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            fp32 var_3879_epsilon_0 = const()[name = string("op_3879_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3879_cast_fp16 = rsqrt(epsilon = var_3879_epsilon_0, x = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3880_cast_fp16 = mul(x = x_95_cast_fp16, y = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427866432)))];
            tensor<fp16, [1, 16, 1024]> input_75_cast_fp16 = mul(x = var_3880_cast_fp16, y = model_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427868544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431014336))))[name = string("model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3885_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_3885_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431020544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434166336))))[name = string("model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_79_cast_fp16 = mul(x = var_3885_cast_fp16, y = linear_54_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434172544))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437318336))))[name = string("model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_97_cast_fp16 = add(x = x_95_cast_fp16, y = linear_55_cast_fp16)[name = string("x_97_cast_fp16")];
            fp16 var_3899_promoted_to_fp16 = const()[name = string("op_3899_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3901_cast_fp16 = pow(x = x_97_cast_fp16, y = var_3899_promoted_to_fp16)[name = string("op_3901_cast_fp16")];
            tensor<int32, [1]> var_3903_axes_0 = const()[name = string("op_3903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3903_keep_dims_0 = const()[name = string("op_3903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3903_cast_fp16 = reduce_mean(axes = var_3903_axes_0, keep_dims = var_3903_keep_dims_0, x = var_3901_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3905_cast_fp16 = add(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            fp32 var_3906_epsilon_0 = const()[name = string("op_3906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3906_cast_fp16 = rsqrt(epsilon = var_3906_epsilon_0, x = var_3905_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3907_cast_fp16 = mul(x = x_97_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437320448)))];
            tensor<fp16, [1, 16, 1024]> input_81_cast_fp16 = mul(x = var_3907_cast_fp16, y = model_layers_8_input_layernorm_weight_to_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437322560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439419776))))[name = string("model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_3916 = const()[name = string("op_3916"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3917_cast_fp16 = reshape(shape = var_3916, x = linear_56_cast_fp16)[name = string("op_3917_cast_fp16")];
            tensor<int32, [4]> x_99_perm_0 = const()[name = string("x_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3925_promoted_to_fp16 = const()[name = string("op_3925_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_99_cast_fp16 = transpose(perm = x_99_perm_0, x = var_3917_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_3927_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3925_promoted_to_fp16)[name = string("op_3927_cast_fp16")];
            tensor<int32, [1]> var_3929_axes_0 = const()[name = string("op_3929_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3929_keep_dims_0 = const()[name = string("op_3929_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3929_cast_fp16 = reduce_mean(axes = var_3929_axes_0, keep_dims = var_3929_keep_dims_0, x = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            fp16 var_3930_to_fp16 = const()[name = string("op_3930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3931_cast_fp16 = add(x = var_3929_cast_fp16, y = var_3930_to_fp16)[name = string("op_3931_cast_fp16")];
            fp32 var_3932_epsilon_0 = const()[name = string("op_3932_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3932_cast_fp16 = rsqrt(epsilon = var_3932_epsilon_0, x = var_3931_cast_fp16)[name = string("op_3932_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3933_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3932_cast_fp16)[name = string("op_3933_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439423936)))];
            tensor<fp16, [1, 16, 16, 128]> x_103_cast_fp16 = mul(x = var_3933_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_103_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439424256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440472896))))[name = string("model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3943_cast_fp16 = reshape(shape = var_3942, x = linear_57_cast_fp16)[name = string("op_3943_cast_fp16")];
            tensor<int32, [4]> x_101_perm_0 = const()[name = string("x_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3951_promoted_to_fp16 = const()[name = string("op_3951_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_101_cast_fp16 = transpose(perm = x_101_perm_0, x = var_3943_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_3953_cast_fp16 = pow(x = x_101_cast_fp16, y = var_3951_promoted_to_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]>([-1])];
            bool var_3955_keep_dims_0 = const()[name = string("op_3955_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3955_cast_fp16 = reduce_mean(axes = var_3955_axes_0, keep_dims = var_3955_keep_dims_0, x = var_3953_cast_fp16)[name = string("op_3955_cast_fp16")];
            fp16 var_3956_to_fp16 = const()[name = string("op_3956_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3957_cast_fp16 = add(x = var_3955_cast_fp16, y = var_3956_to_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3958_cast_fp16 = rsqrt(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3959_cast_fp16 = mul(x = x_101_cast_fp16, y = var_3958_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475008)))];
            tensor<fp16, [1, 8, 16, 128]> x_105_cast_fp16 = mul(x = var_3959_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441523968))))[name = string("model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_3968 = const()[name = string("op_3968"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3969_cast_fp16 = reshape(shape = var_3968, x = linear_58_cast_fp16)[name = string("op_3969_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_33_begin_0 = const()[name = string("first_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_33_end_0 = const()[name = string("first_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_33_end_mask_0 = const()[name = string("first_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_33_cast_fp16 = slice_by_index(begin = first_33_begin_0, end = first_33_end_0, end_mask = first_33_end_mask_0, x = x_103_cast_fp16)[name = string("first_33_cast_fp16")];
            tensor<int32, [4]> second_33_begin_0 = const()[name = string("second_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_33_end_0 = const()[name = string("second_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_33_end_mask_0 = const()[name = string("second_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_33_cast_fp16 = slice_by_index(begin = second_33_begin_0, end = second_33_end_0, end_mask = second_33_end_mask_0, x = x_103_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3983_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3984_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3986_cast_fp16 = sub(x = var_3983_cast_fp16, y = var_3984_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3987_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3988_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3990_cast_fp16 = add(x = var_3987_cast_fp16, y = var_3988_cast_fp16)[name = string("op_3990_cast_fp16")];
            int32 var_3992 = const()[name = string("op_3992"), val = int32(-1)];
            bool q_17_interleave_0 = const()[name = string("q_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_17_cast_fp16 = concat(axis = var_3992, interleave = q_17_interleave_0, values = (var_3986_cast_fp16, var_3990_cast_fp16))[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> first_35_begin_0 = const()[name = string("first_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_35_end_0 = const()[name = string("first_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_35_end_mask_0 = const()[name = string("first_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_35_cast_fp16 = slice_by_index(begin = first_35_begin_0, end = first_35_end_0, end_mask = first_35_end_mask_0, x = x_105_cast_fp16)[name = string("first_35_cast_fp16")];
            tensor<int32, [4]> second_35_begin_0 = const()[name = string("second_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_35_end_0 = const()[name = string("second_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_35_end_mask_0 = const()[name = string("second_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_35_cast_fp16 = slice_by_index(begin = second_35_begin_0, end = second_35_end_0, end_mask = second_35_end_mask_0, x = x_105_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4004_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4004_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4005_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4005_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4007_cast_fp16 = sub(x = var_4004_cast_fp16, y = var_4005_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4008_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4009_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4011_cast_fp16 = add(x = var_4008_cast_fp16, y = var_4009_cast_fp16)[name = string("op_4011_cast_fp16")];
            int32 var_4013 = const()[name = string("op_4013"), val = int32(-1)];
            bool k_17_interleave_0 = const()[name = string("k_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_17_cast_fp16 = concat(axis = var_4013, interleave = k_17_interleave_0, values = (var_4007_cast_fp16, var_4011_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4024_begin_0 = const()[name = string("op_4024_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4024_end_0 = const()[name = string("op_4024_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4024_stride_0 = const()[name = string("op_4024_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4024_cast_fp16 = slice_by_index(begin = var_4024_begin_0, end = var_4024_end_0, end_mask = var_4024_end_mask_0, stride = var_4024_stride_0, x = q_17_cast_fp16)[name = string("op_4024_cast_fp16")];
            tensor<int32, [4]> var_4034_begin_0 = const()[name = string("op_4034_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4034_end_0 = const()[name = string("op_4034_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4034_end_mask_0 = const()[name = string("op_4034_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4034_stride_0 = const()[name = string("op_4034_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4034_cast_fp16 = slice_by_index(begin = var_4034_begin_0, end = var_4034_end_0, end_mask = var_4034_end_mask_0, stride = var_4034_stride_0, x = q_17_cast_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4036 = const()[name = string("op_4036"), val = int32(2)];
            bool stacked_17_interleave_0 = const()[name = string("stacked_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_17_cast_fp16 = concat(axis = var_4036, interleave = stacked_17_interleave_0, values = (var_4024_cast_fp16, var_4034_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4042_begin_0 = const()[name = string("op_4042_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4042_end_0 = const()[name = string("op_4042_end_0"), val = tensor<int32, [4]>([9, 8, 2064, 128])];
            tensor<bool, [4]> var_4042_end_mask_0 = const()[name = string("op_4042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4042_cast_fp16 = slice_by_index(begin = var_4042_begin_0, end = var_4042_end_0, end_mask = var_4042_end_mask_0, x = read_state_0)[name = string("op_4042_cast_fp16")];
            tensor<int32, [4]> var_4052_begin_0 = const()[name = string("op_4052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4052_end_0 = const()[name = string("op_4052_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4052_end_mask_0 = const()[name = string("op_4052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4052_cast_fp16 = slice_by_index(begin = var_4052_begin_0, end = var_4052_end_0, end_mask = var_4052_end_mask_0, x = var_4042_cast_fp16)[name = string("op_4052_cast_fp16")];
            bool var_4061_transpose_x_1 = const()[name = string("op_4061_transpose_x_1"), val = bool(false)];
            bool var_4061_transpose_y_1 = const()[name = string("op_4061_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4061_cast_fp16 = matmul(transpose_x = var_4061_transpose_x_1, transpose_y = var_4061_transpose_y_1, x = var_4052_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4061_cast_fp16")];
            tensor<int32, [4]> var_4064_perm_0 = const()[name = string("op_4064_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4065_to_fp16 = const()[name = string("op_4065_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4064_cast_fp16 = transpose(perm = var_4064_perm_0, x = var_4061_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 2048]> var_4066_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_17_cast_fp16 = add(x = var_4066_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4072_transpose_x_1 = const()[name = string("op_4072_transpose_x_1"), val = bool(false)];
            bool var_4072_transpose_y_1 = const()[name = string("op_4072_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4072_cast_fp16 = matmul(transpose_x = var_4072_transpose_x_1, transpose_y = var_4072_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4072_cast_fp16")];
            fp16 var_4073_to_fp16 = const()[name = string("op_4073_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4074_cast_fp16 = mul(x = var_4072_cast_fp16, y = var_4073_to_fp16)[name = string("op_4074_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4074_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            bool var_4079_interleave_0 = const()[name = string("op_4079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4079_cast_fp16 = concat(axis = var_4078, interleave = var_4079_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4079_cast_fp16")];
            int32 var_4080 = const()[name = string("op_4080"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_17_cast_fp16 = softmax(axis = var_4080, x = var_4079_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4087_begin_0 = const()[name = string("op_4087_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4087_end_0 = const()[name = string("op_4087_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4087_end_mask_0 = const()[name = string("op_4087_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4087_cast_fp16 = slice_by_index(begin = var_4087_begin_0, end = var_4087_end_0, end_mask = var_4087_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4087_cast_fp16")];
            tensor<int32, [4]> var_4092_begin_0 = const()[name = string("op_4092_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4092_end_0 = const()[name = string("op_4092_end_0"), val = tensor<int32, [4]>([9, 8, 2064, 128])];
            tensor<bool, [4]> var_4092_end_mask_0 = const()[name = string("op_4092_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4092_cast_fp16 = slice_by_index(begin = var_4092_begin_0, end = var_4092_end_0, end_mask = var_4092_end_mask_0, x = read_state_1)[name = string("op_4092_cast_fp16")];
            tensor<int32, [4]> var_4102_begin_0 = const()[name = string("op_4102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4102_end_0 = const()[name = string("op_4102_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4102_end_mask_0 = const()[name = string("op_4102_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4102_cast_fp16 = slice_by_index(begin = var_4102_begin_0, end = var_4102_end_0, end_mask = var_4102_end_mask_0, x = var_4092_cast_fp16)[name = string("op_4102_cast_fp16")];
            bool var_4108_transpose_x_0 = const()[name = string("op_4108_transpose_x_0"), val = bool(false)];
            bool var_4108_transpose_y_0 = const()[name = string("op_4108_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4108_cast_fp16 = matmul(transpose_x = var_4108_transpose_x_0, transpose_y = var_4108_transpose_y_0, x = var_4087_cast_fp16, y = var_4102_cast_fp16)[name = string("op_4108_cast_fp16")];
            tensor<int32, [4]> var_4113_begin_0 = const()[name = string("op_4113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4113_end_0 = const()[name = string("op_4113_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4113_end_mask_0 = const()[name = string("op_4113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4113_cast_fp16 = slice_by_index(begin = var_4113_begin_0, end = var_4113_end_0, end_mask = var_4113_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4113_cast_fp16")];
            bool var_4114_transpose_x_0 = const()[name = string("op_4114_transpose_x_0"), val = bool(false)];
            bool var_4114_transpose_y_0 = const()[name = string("op_4114_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_3969_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4114_cast_fp16 = matmul(transpose_x = var_4114_transpose_x_0, transpose_y = var_4114_transpose_y_0, x = var_4113_cast_fp16, y = v_17_cast_fp16)[name = string("op_4114_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4108_cast_fp16, y = var_4114_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4131_begin_0 = const()[name = string("op_4131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4131_end_0 = const()[name = string("op_4131_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4131_end_mask_0 = const()[name = string("op_4131_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4131_cast_fp16 = slice_by_index(begin = var_4131_begin_0, end = var_4131_end_0, end_mask = var_4131_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<int32, [4]> var_4146_begin_0 = const()[name = string("op_4146_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4146_end_0 = const()[name = string("op_4146_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4146_end_mask_0 = const()[name = string("op_4146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4146_cast_fp16 = slice_by_index(begin = var_4146_begin_0, end = var_4146_end_0, end_mask = var_4146_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4146_cast_fp16")];
            int32 var_4148 = const()[name = string("op_4148"), val = int32(1)];
            bool mixed_35_interleave_0 = const()[name = string("mixed_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_35_cast_fp16 = concat(axis = var_4148, interleave = mixed_35_interleave_0, values = (var_4131_cast_fp16, var_4146_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4152_perm_0 = const()[name = string("op_4152_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4156 = const()[name = string("op_4156"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4152_cast_fp16 = transpose(perm = var_4152_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4156, x = var_4152_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441526080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443623296))))[name = string("model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_107_cast_fp16 = add(x = x_97_cast_fp16, y = linear_59_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_4167_promoted_to_fp16 = const()[name = string("op_4167_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4169_cast_fp16 = pow(x = x_107_cast_fp16, y = var_4167_promoted_to_fp16)[name = string("op_4169_cast_fp16")];
            tensor<int32, [1]> var_4171_axes_0 = const()[name = string("op_4171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4171_keep_dims_0 = const()[name = string("op_4171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4171_cast_fp16 = reduce_mean(axes = var_4171_axes_0, keep_dims = var_4171_keep_dims_0, x = var_4169_cast_fp16)[name = string("op_4171_cast_fp16")];
            fp16 var_4172_to_fp16 = const()[name = string("op_4172_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4173_cast_fp16 = add(x = var_4171_cast_fp16, y = var_4172_to_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4174_cast_fp16 = rsqrt(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4175_cast_fp16 = mul(x = x_107_cast_fp16, y = var_4174_cast_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443625408)))];
            tensor<fp16, [1, 16, 1024]> input_85_cast_fp16 = mul(x = var_4175_cast_fp16, y = model_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443627520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446773312))))[name = string("model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4180_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446779520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449925312))))[name = string("model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_89_cast_fp16 = mul(x = var_4180_cast_fp16, y = linear_61_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449931520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453077312))))[name = string("model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = linear_62_cast_fp16)[name = string("x_109_cast_fp16")];
            fp16 var_4194_promoted_to_fp16 = const()[name = string("op_4194_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4196_cast_fp16 = pow(x = x_109_cast_fp16, y = var_4194_promoted_to_fp16)[name = string("op_4196_cast_fp16")];
            tensor<int32, [1]> var_4198_axes_0 = const()[name = string("op_4198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4198_keep_dims_0 = const()[name = string("op_4198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4198_cast_fp16 = reduce_mean(axes = var_4198_axes_0, keep_dims = var_4198_keep_dims_0, x = var_4196_cast_fp16)[name = string("op_4198_cast_fp16")];
            fp16 var_4199_to_fp16 = const()[name = string("op_4199_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4200_cast_fp16 = add(x = var_4198_cast_fp16, y = var_4199_to_fp16)[name = string("op_4200_cast_fp16")];
            fp32 var_4201_epsilon_0 = const()[name = string("op_4201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4201_cast_fp16 = rsqrt(epsilon = var_4201_epsilon_0, x = var_4200_cast_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4202_cast_fp16 = mul(x = x_109_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453079424)))];
            tensor<fp16, [1, 16, 1024]> input_91_cast_fp16 = mul(x = var_4202_cast_fp16, y = model_layers_9_input_layernorm_weight_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453081536))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455178752))))[name = string("model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_4211 = const()[name = string("op_4211"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4212_cast_fp16 = reshape(shape = var_4211, x = linear_63_cast_fp16)[name = string("op_4212_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4220_promoted_to_fp16 = const()[name = string("op_4220_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4212_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4222_cast_fp16 = pow(x = x_111_cast_fp16, y = var_4220_promoted_to_fp16)[name = string("op_4222_cast_fp16")];
            tensor<int32, [1]> var_4224_axes_0 = const()[name = string("op_4224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4224_keep_dims_0 = const()[name = string("op_4224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4224_cast_fp16 = reduce_mean(axes = var_4224_axes_0, keep_dims = var_4224_keep_dims_0, x = var_4222_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp16 var_4225_to_fp16 = const()[name = string("op_4225_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4226_cast_fp16 = add(x = var_4224_cast_fp16, y = var_4225_to_fp16)[name = string("op_4226_cast_fp16")];
            fp32 var_4227_epsilon_0 = const()[name = string("op_4227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4227_cast_fp16 = rsqrt(epsilon = var_4227_epsilon_0, x = var_4226_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4228_cast_fp16 = mul(x = x_111_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455182912)))];
            tensor<fp16, [1, 16, 16, 128]> x_115_cast_fp16 = mul(x = var_4228_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455183232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456231872))))[name = string("model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4237 = const()[name = string("op_4237"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4238_cast_fp16 = reshape(shape = var_4237, x = linear_64_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4246_promoted_to_fp16 = const()[name = string("op_4246_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4238_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4248_cast_fp16 = pow(x = x_113_cast_fp16, y = var_4246_promoted_to_fp16)[name = string("op_4248_cast_fp16")];
            tensor<int32, [1]> var_4250_axes_0 = const()[name = string("op_4250_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4250_keep_dims_0 = const()[name = string("op_4250_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4250_cast_fp16 = reduce_mean(axes = var_4250_axes_0, keep_dims = var_4250_keep_dims_0, x = var_4248_cast_fp16)[name = string("op_4250_cast_fp16")];
            fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4252_cast_fp16 = add(x = var_4250_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")];
            fp32 var_4253_epsilon_0 = const()[name = string("op_4253_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4253_cast_fp16 = rsqrt(epsilon = var_4253_epsilon_0, x = var_4252_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4254_cast_fp16 = mul(x = x_113_cast_fp16, y = var_4253_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456233984)))];
            tensor<fp16, [1, 8, 16, 128]> x_117_cast_fp16 = mul(x = var_4254_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456234304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457282944))))[name = string("model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4263 = const()[name = string("op_4263"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4264_cast_fp16 = reshape(shape = var_4263, x = linear_65_cast_fp16)[name = string("op_4264_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_37_begin_0 = const()[name = string("first_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_37_end_0 = const()[name = string("first_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_37_end_mask_0 = const()[name = string("first_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_37_cast_fp16 = slice_by_index(begin = first_37_begin_0, end = first_37_end_0, end_mask = first_37_end_mask_0, x = x_115_cast_fp16)[name = string("first_37_cast_fp16")];
            tensor<int32, [4]> second_37_begin_0 = const()[name = string("second_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_37_end_0 = const()[name = string("second_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_37_end_mask_0 = const()[name = string("second_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_37_cast_fp16 = slice_by_index(begin = second_37_begin_0, end = second_37_end_0, end_mask = second_37_end_mask_0, x = x_115_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4278_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4278_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4279_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4279_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4281_cast_fp16 = sub(x = var_4278_cast_fp16, y = var_4279_cast_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4282_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4283_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4283_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4285_cast_fp16 = add(x = var_4282_cast_fp16, y = var_4283_cast_fp16)[name = string("op_4285_cast_fp16")];
            int32 var_4287 = const()[name = string("op_4287"), val = int32(-1)];
            bool q_19_interleave_0 = const()[name = string("q_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_19_cast_fp16 = concat(axis = var_4287, interleave = q_19_interleave_0, values = (var_4281_cast_fp16, var_4285_cast_fp16))[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> first_39_begin_0 = const()[name = string("first_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_39_end_0 = const()[name = string("first_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_39_end_mask_0 = const()[name = string("first_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_39_cast_fp16 = slice_by_index(begin = first_39_begin_0, end = first_39_end_0, end_mask = first_39_end_mask_0, x = x_117_cast_fp16)[name = string("first_39_cast_fp16")];
            tensor<int32, [4]> second_39_begin_0 = const()[name = string("second_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_39_end_0 = const()[name = string("second_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_39_end_mask_0 = const()[name = string("second_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_39_cast_fp16 = slice_by_index(begin = second_39_begin_0, end = second_39_end_0, end_mask = second_39_end_mask_0, x = x_117_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4299_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4299_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4300_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4302_cast_fp16 = sub(x = var_4299_cast_fp16, y = var_4300_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4303_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4303_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4304_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4304_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4306_cast_fp16 = add(x = var_4303_cast_fp16, y = var_4304_cast_fp16)[name = string("op_4306_cast_fp16")];
            int32 var_4308 = const()[name = string("op_4308"), val = int32(-1)];
            bool k_19_interleave_0 = const()[name = string("k_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_19_cast_fp16 = concat(axis = var_4308, interleave = k_19_interleave_0, values = (var_4302_cast_fp16, var_4306_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4319_begin_0 = const()[name = string("op_4319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4319_end_0 = const()[name = string("op_4319_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4319_end_mask_0 = const()[name = string("op_4319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4319_stride_0 = const()[name = string("op_4319_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4319_cast_fp16 = slice_by_index(begin = var_4319_begin_0, end = var_4319_end_0, end_mask = var_4319_end_mask_0, stride = var_4319_stride_0, x = q_19_cast_fp16)[name = string("op_4319_cast_fp16")];
            tensor<int32, [4]> var_4329_begin_0 = const()[name = string("op_4329_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4329_end_0 = const()[name = string("op_4329_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4329_end_mask_0 = const()[name = string("op_4329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4329_stride_0 = const()[name = string("op_4329_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4329_cast_fp16 = slice_by_index(begin = var_4329_begin_0, end = var_4329_end_0, end_mask = var_4329_end_mask_0, stride = var_4329_stride_0, x = q_19_cast_fp16)[name = string("op_4329_cast_fp16")];
            int32 var_4331 = const()[name = string("op_4331"), val = int32(2)];
            bool stacked_19_interleave_0 = const()[name = string("stacked_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_19_cast_fp16 = concat(axis = var_4331, interleave = stacked_19_interleave_0, values = (var_4319_cast_fp16, var_4329_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([10, 8, 2064, 128])];
            tensor<bool, [4]> var_4337_end_mask_0 = const()[name = string("op_4337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4337_cast_fp16 = slice_by_index(begin = var_4337_begin_0, end = var_4337_end_0, end_mask = var_4337_end_mask_0, x = read_state_0)[name = string("op_4337_cast_fp16")];
            tensor<int32, [4]> var_4347_begin_0 = const()[name = string("op_4347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4347_end_0 = const()[name = string("op_4347_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4347_end_mask_0 = const()[name = string("op_4347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4347_cast_fp16 = slice_by_index(begin = var_4347_begin_0, end = var_4347_end_0, end_mask = var_4347_end_mask_0, x = var_4337_cast_fp16)[name = string("op_4347_cast_fp16")];
            bool var_4356_transpose_x_1 = const()[name = string("op_4356_transpose_x_1"), val = bool(false)];
            bool var_4356_transpose_y_1 = const()[name = string("op_4356_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4356_cast_fp16 = matmul(transpose_x = var_4356_transpose_x_1, transpose_y = var_4356_transpose_y_1, x = var_4347_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4356_cast_fp16")];
            tensor<int32, [4]> var_4359_perm_0 = const()[name = string("op_4359_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4360_to_fp16 = const()[name = string("op_4360_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4359_cast_fp16 = transpose(perm = var_4359_perm_0, x = var_4356_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 2048]> var_4361_cast_fp16 = mul(x = var_4359_cast_fp16, y = var_4360_to_fp16)[name = string("op_4361_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_19_cast_fp16 = add(x = var_4361_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4367_transpose_x_1 = const()[name = string("op_4367_transpose_x_1"), val = bool(false)];
            bool var_4367_transpose_y_1 = const()[name = string("op_4367_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4367_cast_fp16 = matmul(transpose_x = var_4367_transpose_x_1, transpose_y = var_4367_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4367_cast_fp16")];
            fp16 var_4368_to_fp16 = const()[name = string("op_4368_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4369_cast_fp16 = mul(x = var_4367_cast_fp16, y = var_4368_to_fp16)[name = string("op_4369_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4369_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4373 = const()[name = string("op_4373"), val = int32(-1)];
            bool var_4374_interleave_0 = const()[name = string("op_4374_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4374_cast_fp16 = concat(axis = var_4373, interleave = var_4374_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4374_cast_fp16")];
            int32 var_4375 = const()[name = string("op_4375"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_19_cast_fp16 = softmax(axis = var_4375, x = var_4374_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4382_begin_0 = const()[name = string("op_4382_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4382_end_0 = const()[name = string("op_4382_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4382_end_mask_0 = const()[name = string("op_4382_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4382_cast_fp16 = slice_by_index(begin = var_4382_begin_0, end = var_4382_end_0, end_mask = var_4382_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4382_cast_fp16")];
            tensor<int32, [4]> var_4387_begin_0 = const()[name = string("op_4387_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4387_end_0 = const()[name = string("op_4387_end_0"), val = tensor<int32, [4]>([10, 8, 2064, 128])];
            tensor<bool, [4]> var_4387_end_mask_0 = const()[name = string("op_4387_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4387_cast_fp16 = slice_by_index(begin = var_4387_begin_0, end = var_4387_end_0, end_mask = var_4387_end_mask_0, x = read_state_1)[name = string("op_4387_cast_fp16")];
            tensor<int32, [4]> var_4397_begin_0 = const()[name = string("op_4397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4397_end_0 = const()[name = string("op_4397_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4397_end_mask_0 = const()[name = string("op_4397_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4397_cast_fp16 = slice_by_index(begin = var_4397_begin_0, end = var_4397_end_0, end_mask = var_4397_end_mask_0, x = var_4387_cast_fp16)[name = string("op_4397_cast_fp16")];
            bool var_4403_transpose_x_0 = const()[name = string("op_4403_transpose_x_0"), val = bool(false)];
            bool var_4403_transpose_y_0 = const()[name = string("op_4403_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4403_cast_fp16 = matmul(transpose_x = var_4403_transpose_x_0, transpose_y = var_4403_transpose_y_0, x = var_4382_cast_fp16, y = var_4397_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<int32, [4]> var_4408_begin_0 = const()[name = string("op_4408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4408_end_0 = const()[name = string("op_4408_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4408_end_mask_0 = const()[name = string("op_4408_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4408_cast_fp16")];
            bool var_4409_transpose_x_0 = const()[name = string("op_4409_transpose_x_0"), val = bool(false)];
            bool var_4409_transpose_y_0 = const()[name = string("op_4409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4264_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4409_cast_fp16 = matmul(transpose_x = var_4409_transpose_x_0, transpose_y = var_4409_transpose_y_0, x = var_4408_cast_fp16, y = v_19_cast_fp16)[name = string("op_4409_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4403_cast_fp16, y = var_4409_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4426_begin_0 = const()[name = string("op_4426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4426_end_0 = const()[name = string("op_4426_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4426_end_mask_0 = const()[name = string("op_4426_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4426_cast_fp16 = slice_by_index(begin = var_4426_begin_0, end = var_4426_end_0, end_mask = var_4426_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<int32, [4]> var_4441_begin_0 = const()[name = string("op_4441_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4441_end_0 = const()[name = string("op_4441_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4441_end_mask_0 = const()[name = string("op_4441_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4441_cast_fp16 = slice_by_index(begin = var_4441_begin_0, end = var_4441_end_0, end_mask = var_4441_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4441_cast_fp16")];
            int32 var_4443 = const()[name = string("op_4443"), val = int32(1)];
            bool mixed_39_interleave_0 = const()[name = string("mixed_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_39_cast_fp16 = concat(axis = var_4443, interleave = mixed_39_interleave_0, values = (var_4426_cast_fp16, var_4441_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4447_perm_0 = const()[name = string("op_4447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4451 = const()[name = string("op_4451"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4447_cast_fp16 = transpose(perm = var_4447_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4451, x = var_4447_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457285056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459382272))))[name = string("model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = linear_66_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_4462_promoted_to_fp16 = const()[name = string("op_4462_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4464_cast_fp16 = pow(x = x_119_cast_fp16, y = var_4462_promoted_to_fp16)[name = string("op_4464_cast_fp16")];
            tensor<int32, [1]> var_4466_axes_0 = const()[name = string("op_4466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4466_keep_dims_0 = const()[name = string("op_4466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4466_cast_fp16 = reduce_mean(axes = var_4466_axes_0, keep_dims = var_4466_keep_dims_0, x = var_4464_cast_fp16)[name = string("op_4466_cast_fp16")];
            fp16 var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4468_cast_fp16 = add(x = var_4466_cast_fp16, y = var_4467_to_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4470_cast_fp16 = mul(x = x_119_cast_fp16, y = var_4469_cast_fp16)[name = string("op_4470_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459384384)))];
            tensor<fp16, [1, 16, 1024]> input_95_cast_fp16 = mul(x = var_4470_cast_fp16, y = model_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459386496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462532288))))[name = string("model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4475_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462538496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465684288))))[name = string("model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_99_cast_fp16 = mul(x = var_4475_cast_fp16, y = linear_68_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465690496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468836288))))[name = string("model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_121_cast_fp16 = add(x = x_119_cast_fp16, y = linear_69_cast_fp16)[name = string("x_121_cast_fp16")];
            fp16 var_4489_promoted_to_fp16 = const()[name = string("op_4489_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4491_cast_fp16 = pow(x = x_121_cast_fp16, y = var_4489_promoted_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4493_keep_dims_0 = const()[name = string("op_4493_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4493_cast_fp16 = reduce_mean(axes = var_4493_axes_0, keep_dims = var_4493_keep_dims_0, x = var_4491_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4495_cast_fp16 = add(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("op_4495_cast_fp16")];
            fp32 var_4496_epsilon_0 = const()[name = string("op_4496_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4496_cast_fp16 = rsqrt(epsilon = var_4496_epsilon_0, x = var_4495_cast_fp16)[name = string("op_4496_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4497_cast_fp16 = mul(x = x_121_cast_fp16, y = var_4496_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468838400)))];
            tensor<fp16, [1, 16, 1024]> input_101_cast_fp16 = mul(x = var_4497_cast_fp16, y = model_layers_10_input_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468840512))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470937728))))[name = string("model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4507_cast_fp16 = reshape(shape = var_4506, x = linear_70_cast_fp16)[name = string("op_4507_cast_fp16")];
            tensor<int32, [4]> x_123_perm_0 = const()[name = string("x_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4515_promoted_to_fp16 = const()[name = string("op_4515_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_123_cast_fp16 = transpose(perm = x_123_perm_0, x = var_4507_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4517_cast_fp16 = pow(x = x_123_cast_fp16, y = var_4515_promoted_to_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4519_axes_0 = const()[name = string("op_4519_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4519_keep_dims_0 = const()[name = string("op_4519_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4519_cast_fp16 = reduce_mean(axes = var_4519_axes_0, keep_dims = var_4519_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4519_cast_fp16")];
            fp16 var_4520_to_fp16 = const()[name = string("op_4520_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4521_cast_fp16 = add(x = var_4519_cast_fp16, y = var_4520_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4522_cast_fp16 = rsqrt(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4523_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4522_cast_fp16)[name = string("op_4523_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470941888)))];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = mul(x = var_4523_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470942208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471990848))))[name = string("model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4532 = const()[name = string("op_4532"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4533_cast_fp16 = reshape(shape = var_4532, x = linear_71_cast_fp16)[name = string("op_4533_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4541_promoted_to_fp16 = const()[name = string("op_4541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_4533_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4543_cast_fp16 = pow(x = x_125_cast_fp16, y = var_4541_promoted_to_fp16)[name = string("op_4543_cast_fp16")];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4545_keep_dims_0 = const()[name = string("op_4545_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4545_cast_fp16 = reduce_mean(axes = var_4545_axes_0, keep_dims = var_4545_keep_dims_0, x = var_4543_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4547_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("op_4547_cast_fp16")];
            fp32 var_4548_epsilon_0 = const()[name = string("op_4548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4548_cast_fp16 = rsqrt(epsilon = var_4548_epsilon_0, x = var_4547_cast_fp16)[name = string("op_4548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4549_cast_fp16 = mul(x = x_125_cast_fp16, y = var_4548_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471992960)))];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = mul(x = var_4549_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471993280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473041920))))[name = string("model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4558 = const()[name = string("op_4558"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4559_cast_fp16 = reshape(shape = var_4558, x = linear_72_cast_fp16)[name = string("op_4559_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_41_begin_0 = const()[name = string("first_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_41_end_0 = const()[name = string("first_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_41_end_mask_0 = const()[name = string("first_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_41_cast_fp16 = slice_by_index(begin = first_41_begin_0, end = first_41_end_0, end_mask = first_41_end_mask_0, x = x_127_cast_fp16)[name = string("first_41_cast_fp16")];
            tensor<int32, [4]> second_41_begin_0 = const()[name = string("second_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_41_end_0 = const()[name = string("second_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_41_end_mask_0 = const()[name = string("second_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_41_cast_fp16 = slice_by_index(begin = second_41_begin_0, end = second_41_end_0, end_mask = second_41_end_mask_0, x = x_127_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4573_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4574_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4574_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4576_cast_fp16 = sub(x = var_4573_cast_fp16, y = var_4574_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4577_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4578_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4578_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4580_cast_fp16 = add(x = var_4577_cast_fp16, y = var_4578_cast_fp16)[name = string("op_4580_cast_fp16")];
            int32 var_4582 = const()[name = string("op_4582"), val = int32(-1)];
            bool q_21_interleave_0 = const()[name = string("q_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_21_cast_fp16 = concat(axis = var_4582, interleave = q_21_interleave_0, values = (var_4576_cast_fp16, var_4580_cast_fp16))[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> first_43_begin_0 = const()[name = string("first_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_43_end_0 = const()[name = string("first_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_43_end_mask_0 = const()[name = string("first_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_43_cast_fp16 = slice_by_index(begin = first_43_begin_0, end = first_43_end_0, end_mask = first_43_end_mask_0, x = x_129_cast_fp16)[name = string("first_43_cast_fp16")];
            tensor<int32, [4]> second_43_begin_0 = const()[name = string("second_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_43_end_0 = const()[name = string("second_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_43_end_mask_0 = const()[name = string("second_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_43_cast_fp16 = slice_by_index(begin = second_43_begin_0, end = second_43_end_0, end_mask = second_43_end_mask_0, x = x_129_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4594_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4594_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4595_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4595_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4597_cast_fp16 = sub(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4598_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4599_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4599_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4601_cast_fp16 = add(x = var_4598_cast_fp16, y = var_4599_cast_fp16)[name = string("op_4601_cast_fp16")];
            int32 var_4603 = const()[name = string("op_4603"), val = int32(-1)];
            bool k_21_interleave_0 = const()[name = string("k_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_21_cast_fp16 = concat(axis = var_4603, interleave = k_21_interleave_0, values = (var_4597_cast_fp16, var_4601_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4614_begin_0 = const()[name = string("op_4614_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4614_end_0 = const()[name = string("op_4614_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4614_end_mask_0 = const()[name = string("op_4614_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4614_stride_0 = const()[name = string("op_4614_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4614_cast_fp16 = slice_by_index(begin = var_4614_begin_0, end = var_4614_end_0, end_mask = var_4614_end_mask_0, stride = var_4614_stride_0, x = q_21_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<int32, [4]> var_4624_begin_0 = const()[name = string("op_4624_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4624_end_0 = const()[name = string("op_4624_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4624_end_mask_0 = const()[name = string("op_4624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4624_stride_0 = const()[name = string("op_4624_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4624_cast_fp16 = slice_by_index(begin = var_4624_begin_0, end = var_4624_end_0, end_mask = var_4624_end_mask_0, stride = var_4624_stride_0, x = q_21_cast_fp16)[name = string("op_4624_cast_fp16")];
            int32 var_4626 = const()[name = string("op_4626"), val = int32(2)];
            bool stacked_21_interleave_0 = const()[name = string("stacked_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_21_cast_fp16 = concat(axis = var_4626, interleave = stacked_21_interleave_0, values = (var_4614_cast_fp16, var_4624_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4632_begin_0 = const()[name = string("op_4632_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4632_end_0 = const()[name = string("op_4632_end_0"), val = tensor<int32, [4]>([11, 8, 2064, 128])];
            tensor<bool, [4]> var_4632_end_mask_0 = const()[name = string("op_4632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4632_cast_fp16 = slice_by_index(begin = var_4632_begin_0, end = var_4632_end_0, end_mask = var_4632_end_mask_0, x = read_state_0)[name = string("op_4632_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = var_4632_cast_fp16)[name = string("op_4642_cast_fp16")];
            bool var_4651_transpose_x_1 = const()[name = string("op_4651_transpose_x_1"), val = bool(false)];
            bool var_4651_transpose_y_1 = const()[name = string("op_4651_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_1, transpose_y = var_4651_transpose_y_1, x = var_4642_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [4]> var_4654_perm_0 = const()[name = string("op_4654_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4655_to_fp16 = const()[name = string("op_4655_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4654_cast_fp16 = transpose(perm = var_4654_perm_0, x = var_4651_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 2048]> var_4656_cast_fp16 = mul(x = var_4654_cast_fp16, y = var_4655_to_fp16)[name = string("op_4656_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_21_cast_fp16 = add(x = var_4656_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4662_transpose_x_1 = const()[name = string("op_4662_transpose_x_1"), val = bool(false)];
            bool var_4662_transpose_y_1 = const()[name = string("op_4662_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4662_cast_fp16 = matmul(transpose_x = var_4662_transpose_x_1, transpose_y = var_4662_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4662_cast_fp16")];
            fp16 var_4663_to_fp16 = const()[name = string("op_4663_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4664_cast_fp16 = mul(x = var_4662_cast_fp16, y = var_4663_to_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4664_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4668 = const()[name = string("op_4668"), val = int32(-1)];
            bool var_4669_interleave_0 = const()[name = string("op_4669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4669_cast_fp16 = concat(axis = var_4668, interleave = var_4669_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4669_cast_fp16")];
            int32 var_4670 = const()[name = string("op_4670"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_21_cast_fp16 = softmax(axis = var_4670, x = var_4669_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4677_begin_0 = const()[name = string("op_4677_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4677_end_0 = const()[name = string("op_4677_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4677_end_mask_0 = const()[name = string("op_4677_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4677_cast_fp16 = slice_by_index(begin = var_4677_begin_0, end = var_4677_end_0, end_mask = var_4677_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4677_cast_fp16")];
            tensor<int32, [4]> var_4682_begin_0 = const()[name = string("op_4682_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4682_end_0 = const()[name = string("op_4682_end_0"), val = tensor<int32, [4]>([11, 8, 2064, 128])];
            tensor<bool, [4]> var_4682_end_mask_0 = const()[name = string("op_4682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4682_cast_fp16 = slice_by_index(begin = var_4682_begin_0, end = var_4682_end_0, end_mask = var_4682_end_mask_0, x = read_state_1)[name = string("op_4682_cast_fp16")];
            tensor<int32, [4]> var_4692_begin_0 = const()[name = string("op_4692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4692_end_0 = const()[name = string("op_4692_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4692_end_mask_0 = const()[name = string("op_4692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4692_cast_fp16 = slice_by_index(begin = var_4692_begin_0, end = var_4692_end_0, end_mask = var_4692_end_mask_0, x = var_4682_cast_fp16)[name = string("op_4692_cast_fp16")];
            bool var_4698_transpose_x_0 = const()[name = string("op_4698_transpose_x_0"), val = bool(false)];
            bool var_4698_transpose_y_0 = const()[name = string("op_4698_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4698_cast_fp16 = matmul(transpose_x = var_4698_transpose_x_0, transpose_y = var_4698_transpose_y_0, x = var_4677_cast_fp16, y = var_4692_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<int32, [4]> var_4703_begin_0 = const()[name = string("op_4703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4703_end_0 = const()[name = string("op_4703_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4703_end_mask_0 = const()[name = string("op_4703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4703_cast_fp16 = slice_by_index(begin = var_4703_begin_0, end = var_4703_end_0, end_mask = var_4703_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4703_cast_fp16")];
            bool var_4704_transpose_x_0 = const()[name = string("op_4704_transpose_x_0"), val = bool(false)];
            bool var_4704_transpose_y_0 = const()[name = string("op_4704_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4559_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_4704_cast_fp16 = matmul(transpose_x = var_4704_transpose_x_0, transpose_y = var_4704_transpose_y_0, x = var_4703_cast_fp16, y = v_21_cast_fp16)[name = string("op_4704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4704_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_4721_begin_0 = const()[name = string("op_4721_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4721_end_0 = const()[name = string("op_4721_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4721_end_mask_0 = const()[name = string("op_4721_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4721_cast_fp16 = slice_by_index(begin = var_4721_begin_0, end = var_4721_end_0, end_mask = var_4721_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<int32, [4]> var_4736_begin_0 = const()[name = string("op_4736_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4736_end_0 = const()[name = string("op_4736_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4736_end_mask_0 = const()[name = string("op_4736_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4736_cast_fp16 = slice_by_index(begin = var_4736_begin_0, end = var_4736_end_0, end_mask = var_4736_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4736_cast_fp16")];
            int32 var_4738 = const()[name = string("op_4738"), val = int32(1)];
            bool mixed_43_interleave_0 = const()[name = string("mixed_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_43_cast_fp16 = concat(axis = var_4738, interleave = mixed_43_interleave_0, values = (var_4721_cast_fp16, var_4736_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_4742_perm_0 = const()[name = string("op_4742_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4746 = const()[name = string("op_4746"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4742_cast_fp16 = transpose(perm = var_4742_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_4746, x = var_4742_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473044032))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475141248))))[name = string("model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_131_cast_fp16 = add(x = x_121_cast_fp16, y = linear_73_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_4757_promoted_to_fp16 = const()[name = string("op_4757_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4759_cast_fp16 = pow(x = x_131_cast_fp16, y = var_4757_promoted_to_fp16)[name = string("op_4759_cast_fp16")];
            tensor<int32, [1]> var_4761_axes_0 = const()[name = string("op_4761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4761_keep_dims_0 = const()[name = string("op_4761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4761_cast_fp16 = reduce_mean(axes = var_4761_axes_0, keep_dims = var_4761_keep_dims_0, x = var_4759_cast_fp16)[name = string("op_4761_cast_fp16")];
            fp16 var_4762_to_fp16 = const()[name = string("op_4762_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4763_cast_fp16 = add(x = var_4761_cast_fp16, y = var_4762_to_fp16)[name = string("op_4763_cast_fp16")];
            fp32 var_4764_epsilon_0 = const()[name = string("op_4764_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4764_cast_fp16 = rsqrt(epsilon = var_4764_epsilon_0, x = var_4763_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_4764_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475143360)))];
            tensor<fp16, [1, 16, 1024]> input_105_cast_fp16 = mul(x = var_4765_cast_fp16, y = model_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475145472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478291264))))[name = string("model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4770_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_4770_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478297472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481443264))))[name = string("model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_109_cast_fp16 = mul(x = var_4770_cast_fp16, y = linear_75_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481449472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484595264))))[name = string("model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_133_cast_fp16 = add(x = x_131_cast_fp16, y = linear_76_cast_fp16)[name = string("x_133_cast_fp16")];
            fp16 var_4784_promoted_to_fp16 = const()[name = string("op_4784_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4786_cast_fp16 = pow(x = x_133_cast_fp16, y = var_4784_promoted_to_fp16)[name = string("op_4786_cast_fp16")];
            tensor<int32, [1]> var_4788_axes_0 = const()[name = string("op_4788_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4788_keep_dims_0 = const()[name = string("op_4788_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4788_cast_fp16 = reduce_mean(axes = var_4788_axes_0, keep_dims = var_4788_keep_dims_0, x = var_4786_cast_fp16)[name = string("op_4788_cast_fp16")];
            fp16 var_4789_to_fp16 = const()[name = string("op_4789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4790_cast_fp16 = add(x = var_4788_cast_fp16, y = var_4789_to_fp16)[name = string("op_4790_cast_fp16")];
            fp32 var_4791_epsilon_0 = const()[name = string("op_4791_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4791_cast_fp16 = rsqrt(epsilon = var_4791_epsilon_0, x = var_4790_cast_fp16)[name = string("op_4791_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4792_cast_fp16 = mul(x = x_133_cast_fp16, y = var_4791_cast_fp16)[name = string("op_4792_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484597376)))];
            tensor<fp16, [1, 16, 1024]> input_111_cast_fp16 = mul(x = var_4792_cast_fp16, y = model_layers_11_input_layernorm_weight_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484599488))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486696704))))[name = string("model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_4801 = const()[name = string("op_4801"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4802_cast_fp16 = reshape(shape = var_4801, x = linear_77_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4810_promoted_to_fp16 = const()[name = string("op_4810_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_4802_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_4812_cast_fp16 = pow(x = x_135_cast_fp16, y = var_4810_promoted_to_fp16)[name = string("op_4812_cast_fp16")];
            tensor<int32, [1]> var_4814_axes_0 = const()[name = string("op_4814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4814_keep_dims_0 = const()[name = string("op_4814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = reduce_mean(axes = var_4814_axes_0, keep_dims = var_4814_keep_dims_0, x = var_4812_cast_fp16)[name = string("op_4814_cast_fp16")];
            fp16 var_4815_to_fp16 = const()[name = string("op_4815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = add(x = var_4814_cast_fp16, y = var_4815_to_fp16)[name = string("op_4816_cast_fp16")];
            fp32 var_4817_epsilon_0 = const()[name = string("op_4817_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4817_cast_fp16 = rsqrt(epsilon = var_4817_epsilon_0, x = var_4816_cast_fp16)[name = string("op_4817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4818_cast_fp16 = mul(x = x_135_cast_fp16, y = var_4817_cast_fp16)[name = string("op_4818_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486700864)))];
            tensor<fp16, [1, 16, 16, 128]> x_139_cast_fp16 = mul(x = var_4818_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486701184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487749824))))[name = string("model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_4827 = const()[name = string("op_4827"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4828_cast_fp16 = reshape(shape = var_4827, x = linear_78_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<int32, [4]> x_137_perm_0 = const()[name = string("x_137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4836_promoted_to_fp16 = const()[name = string("op_4836_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_137_cast_fp16 = transpose(perm = x_137_perm_0, x = var_4828_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_4838_cast_fp16 = pow(x = x_137_cast_fp16, y = var_4836_promoted_to_fp16)[name = string("op_4838_cast_fp16")];
            tensor<int32, [1]> var_4840_axes_0 = const()[name = string("op_4840_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4840_keep_dims_0 = const()[name = string("op_4840_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4840_cast_fp16 = reduce_mean(axes = var_4840_axes_0, keep_dims = var_4840_keep_dims_0, x = var_4838_cast_fp16)[name = string("op_4840_cast_fp16")];
            fp16 var_4841_to_fp16 = const()[name = string("op_4841_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4842_cast_fp16 = add(x = var_4840_cast_fp16, y = var_4841_to_fp16)[name = string("op_4842_cast_fp16")];
            fp32 var_4843_epsilon_0 = const()[name = string("op_4843_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = rsqrt(epsilon = var_4843_epsilon_0, x = var_4842_cast_fp16)[name = string("op_4843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4844_cast_fp16 = mul(x = x_137_cast_fp16, y = var_4843_cast_fp16)[name = string("op_4844_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487751936)))];
            tensor<fp16, [1, 8, 16, 128]> x_141_cast_fp16 = mul(x = var_4844_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487752256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488800896))))[name = string("model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_4853 = const()[name = string("op_4853"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4854_cast_fp16 = reshape(shape = var_4853, x = linear_79_cast_fp16)[name = string("op_4854_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_45_begin_0 = const()[name = string("first_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_45_end_0 = const()[name = string("first_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_45_end_mask_0 = const()[name = string("first_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_45_cast_fp16 = slice_by_index(begin = first_45_begin_0, end = first_45_end_0, end_mask = first_45_end_mask_0, x = x_139_cast_fp16)[name = string("first_45_cast_fp16")];
            tensor<int32, [4]> second_45_begin_0 = const()[name = string("second_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_45_end_0 = const()[name = string("second_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_45_end_mask_0 = const()[name = string("second_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_45_cast_fp16 = slice_by_index(begin = second_45_begin_0, end = second_45_end_0, end_mask = second_45_end_mask_0, x = x_139_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4868_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4869_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4869_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4871_cast_fp16 = sub(x = var_4868_cast_fp16, y = var_4869_cast_fp16)[name = string("op_4871_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4872_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4872_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4873_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4873_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = add(x = var_4872_cast_fp16, y = var_4873_cast_fp16)[name = string("op_4875_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-1)];
            bool q_23_interleave_0 = const()[name = string("q_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_23_cast_fp16 = concat(axis = var_4877, interleave = q_23_interleave_0, values = (var_4871_cast_fp16, var_4875_cast_fp16))[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> first_47_begin_0 = const()[name = string("first_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_47_end_0 = const()[name = string("first_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_47_end_mask_0 = const()[name = string("first_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_47_cast_fp16 = slice_by_index(begin = first_47_begin_0, end = first_47_end_0, end_mask = first_47_end_mask_0, x = x_141_cast_fp16)[name = string("first_47_cast_fp16")];
            tensor<int32, [4]> second_47_begin_0 = const()[name = string("second_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_47_end_0 = const()[name = string("second_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_47_end_mask_0 = const()[name = string("second_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_47_cast_fp16 = slice_by_index(begin = second_47_begin_0, end = second_47_end_0, end_mask = second_47_end_mask_0, x = x_141_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4889_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4890_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4892_cast_fp16 = sub(x = var_4889_cast_fp16, y = var_4890_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4893_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4894_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = add(x = var_4893_cast_fp16, y = var_4894_cast_fp16)[name = string("op_4896_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(-1)];
            bool k_23_interleave_0 = const()[name = string("k_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_23_cast_fp16 = concat(axis = var_4898, interleave = k_23_interleave_0, values = (var_4892_cast_fp16, var_4896_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_4909_begin_0 = const()[name = string("op_4909_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4909_end_0 = const()[name = string("op_4909_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4909_end_mask_0 = const()[name = string("op_4909_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4909_stride_0 = const()[name = string("op_4909_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4909_cast_fp16 = slice_by_index(begin = var_4909_begin_0, end = var_4909_end_0, end_mask = var_4909_end_mask_0, stride = var_4909_stride_0, x = q_23_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [4]> var_4919_begin_0 = const()[name = string("op_4919_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4919_end_0 = const()[name = string("op_4919_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4919_end_mask_0 = const()[name = string("op_4919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4919_stride_0 = const()[name = string("op_4919_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4919_cast_fp16 = slice_by_index(begin = var_4919_begin_0, end = var_4919_end_0, end_mask = var_4919_end_mask_0, stride = var_4919_stride_0, x = q_23_cast_fp16)[name = string("op_4919_cast_fp16")];
            int32 var_4921 = const()[name = string("op_4921"), val = int32(2)];
            bool stacked_23_interleave_0 = const()[name = string("stacked_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_23_cast_fp16 = concat(axis = var_4921, interleave = stacked_23_interleave_0, values = (var_4909_cast_fp16, var_4919_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = string("op_4927_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = string("op_4927_end_0"), val = tensor<int32, [4]>([12, 8, 2064, 128])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = string("op_4927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4927_cast_fp16 = slice_by_index(begin = var_4927_begin_0, end = var_4927_end_0, end_mask = var_4927_end_mask_0, x = read_state_0)[name = string("op_4927_cast_fp16")];
            tensor<int32, [4]> var_4937_begin_0 = const()[name = string("op_4937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4937_end_0 = const()[name = string("op_4937_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4937_end_mask_0 = const()[name = string("op_4937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4937_cast_fp16 = slice_by_index(begin = var_4937_begin_0, end = var_4937_end_0, end_mask = var_4937_end_mask_0, x = var_4927_cast_fp16)[name = string("op_4937_cast_fp16")];
            bool var_4946_transpose_x_1 = const()[name = string("op_4946_transpose_x_1"), val = bool(false)];
            bool var_4946_transpose_y_1 = const()[name = string("op_4946_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_4946_cast_fp16 = matmul(transpose_x = var_4946_transpose_x_1, transpose_y = var_4946_transpose_y_1, x = var_4937_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4949_perm_0 = const()[name = string("op_4949_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4950_to_fp16 = const()[name = string("op_4950_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_4949_cast_fp16 = transpose(perm = var_4949_perm_0, x = var_4946_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 2048]> var_4951_cast_fp16 = mul(x = var_4949_cast_fp16, y = var_4950_to_fp16)[name = string("op_4951_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_23_cast_fp16 = add(x = var_4951_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_4957_transpose_x_1 = const()[name = string("op_4957_transpose_x_1"), val = bool(false)];
            bool var_4957_transpose_y_1 = const()[name = string("op_4957_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4957_cast_fp16 = matmul(transpose_x = var_4957_transpose_x_1, transpose_y = var_4957_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_4957_cast_fp16")];
            fp16 var_4958_to_fp16 = const()[name = string("op_4958_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4959_cast_fp16 = mul(x = var_4957_cast_fp16, y = var_4958_to_fp16)[name = string("op_4959_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_4959_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_4963 = const()[name = string("op_4963"), val = int32(-1)];
            bool var_4964_interleave_0 = const()[name = string("op_4964_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_4964_cast_fp16 = concat(axis = var_4963, interleave = var_4964_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_4964_cast_fp16")];
            int32 var_4965 = const()[name = string("op_4965"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_23_cast_fp16 = softmax(axis = var_4965, x = var_4964_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_4972_begin_0 = const()[name = string("op_4972_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4972_end_0 = const()[name = string("op_4972_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_4972_end_mask_0 = const()[name = string("op_4972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_4972_cast_fp16 = slice_by_index(begin = var_4972_begin_0, end = var_4972_end_0, end_mask = var_4972_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<int32, [4]> var_4977_begin_0 = const()[name = string("op_4977_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4977_end_0 = const()[name = string("op_4977_end_0"), val = tensor<int32, [4]>([12, 8, 2064, 128])];
            tensor<bool, [4]> var_4977_end_mask_0 = const()[name = string("op_4977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_4977_cast_fp16 = slice_by_index(begin = var_4977_begin_0, end = var_4977_end_0, end_mask = var_4977_end_mask_0, x = read_state_1)[name = string("op_4977_cast_fp16")];
            tensor<int32, [4]> var_4987_begin_0 = const()[name = string("op_4987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4987_end_0 = const()[name = string("op_4987_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_4987_end_mask_0 = const()[name = string("op_4987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_4987_cast_fp16 = slice_by_index(begin = var_4987_begin_0, end = var_4987_end_0, end_mask = var_4987_end_mask_0, x = var_4977_cast_fp16)[name = string("op_4987_cast_fp16")];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = var_4972_cast_fp16, y = var_4987_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<int32, [4]> var_4998_begin_0 = const()[name = string("op_4998_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_4998_end_0 = const()[name = string("op_4998_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_4998_end_mask_0 = const()[name = string("op_4998_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4998_cast_fp16 = slice_by_index(begin = var_4998_begin_0, end = var_4998_end_0, end_mask = var_4998_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4998_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_4854_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4998_cast_fp16, y = v_23_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_4993_cast_fp16, y = var_4999_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5016_begin_0 = const()[name = string("op_5016_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5016_end_0 = const()[name = string("op_5016_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5016_end_mask_0 = const()[name = string("op_5016_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5016_cast_fp16 = slice_by_index(begin = var_5016_begin_0, end = var_5016_end_0, end_mask = var_5016_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5016_cast_fp16")];
            tensor<int32, [4]> var_5031_begin_0 = const()[name = string("op_5031_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5031_end_0 = const()[name = string("op_5031_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5031_end_mask_0 = const()[name = string("op_5031_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5031_cast_fp16 = slice_by_index(begin = var_5031_begin_0, end = var_5031_end_0, end_mask = var_5031_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5031_cast_fp16")];
            int32 var_5033 = const()[name = string("op_5033"), val = int32(1)];
            bool mixed_47_interleave_0 = const()[name = string("mixed_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_47_cast_fp16 = concat(axis = var_5033, interleave = mixed_47_interleave_0, values = (var_5016_cast_fp16, var_5031_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5037_perm_0 = const()[name = string("op_5037_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5041 = const()[name = string("op_5041"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5037_cast_fp16 = transpose(perm = var_5037_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5041, x = var_5037_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488803008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490900224))))[name = string("model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = add(x = x_133_cast_fp16, y = linear_80_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_5052_promoted_to_fp16 = const()[name = string("op_5052_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5054_cast_fp16 = pow(x = x_143_cast_fp16, y = var_5052_promoted_to_fp16)[name = string("op_5054_cast_fp16")];
            tensor<int32, [1]> var_5056_axes_0 = const()[name = string("op_5056_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5056_keep_dims_0 = const()[name = string("op_5056_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5056_cast_fp16 = reduce_mean(axes = var_5056_axes_0, keep_dims = var_5056_keep_dims_0, x = var_5054_cast_fp16)[name = string("op_5056_cast_fp16")];
            fp16 var_5057_to_fp16 = const()[name = string("op_5057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5058_cast_fp16 = add(x = var_5056_cast_fp16, y = var_5057_to_fp16)[name = string("op_5058_cast_fp16")];
            fp32 var_5059_epsilon_0 = const()[name = string("op_5059_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5059_cast_fp16 = rsqrt(epsilon = var_5059_epsilon_0, x = var_5058_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = mul(x = x_143_cast_fp16, y = var_5059_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490902336)))];
            tensor<fp16, [1, 16, 1024]> input_115_cast_fp16 = mul(x = var_5060_cast_fp16, y = model_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490904448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494050240))))[name = string("model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5065_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494056448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497202240))))[name = string("model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_119_cast_fp16 = mul(x = var_5065_cast_fp16, y = linear_82_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497208448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500354240))))[name = string("model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_145_cast_fp16 = add(x = x_143_cast_fp16, y = linear_83_cast_fp16)[name = string("x_145_cast_fp16")];
            fp16 var_5079_promoted_to_fp16 = const()[name = string("op_5079_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5081_cast_fp16 = pow(x = x_145_cast_fp16, y = var_5079_promoted_to_fp16)[name = string("op_5081_cast_fp16")];
            tensor<int32, [1]> var_5083_axes_0 = const()[name = string("op_5083_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5083_keep_dims_0 = const()[name = string("op_5083_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5083_cast_fp16 = reduce_mean(axes = var_5083_axes_0, keep_dims = var_5083_keep_dims_0, x = var_5081_cast_fp16)[name = string("op_5083_cast_fp16")];
            fp16 var_5084_to_fp16 = const()[name = string("op_5084_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5085_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5084_to_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5086_cast_fp16 = rsqrt(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5087_cast_fp16 = mul(x = x_145_cast_fp16, y = var_5086_cast_fp16)[name = string("op_5087_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500356352)))];
            tensor<fp16, [1, 16, 1024]> input_121_cast_fp16 = mul(x = var_5087_cast_fp16, y = model_layers_12_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500358464))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502455680))))[name = string("model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5096 = const()[name = string("op_5096"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5097_cast_fp16 = reshape(shape = var_5096, x = linear_84_cast_fp16)[name = string("op_5097_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5105_promoted_to_fp16 = const()[name = string("op_5105_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_5097_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5107_cast_fp16 = pow(x = x_147_cast_fp16, y = var_5105_promoted_to_fp16)[name = string("op_5107_cast_fp16")];
            tensor<int32, [1]> var_5109_axes_0 = const()[name = string("op_5109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5109_keep_dims_0 = const()[name = string("op_5109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5109_cast_fp16 = reduce_mean(axes = var_5109_axes_0, keep_dims = var_5109_keep_dims_0, x = var_5107_cast_fp16)[name = string("op_5109_cast_fp16")];
            fp16 var_5110_to_fp16 = const()[name = string("op_5110_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5111_cast_fp16 = add(x = var_5109_cast_fp16, y = var_5110_to_fp16)[name = string("op_5111_cast_fp16")];
            fp32 var_5112_epsilon_0 = const()[name = string("op_5112_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5112_cast_fp16 = rsqrt(epsilon = var_5112_epsilon_0, x = var_5111_cast_fp16)[name = string("op_5112_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5113_cast_fp16 = mul(x = x_147_cast_fp16, y = var_5112_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502459840)))];
            tensor<fp16, [1, 16, 16, 128]> x_151_cast_fp16 = mul(x = var_5113_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502460160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503508800))))[name = string("model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5122 = const()[name = string("op_5122"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5123_cast_fp16 = reshape(shape = var_5122, x = linear_85_cast_fp16)[name = string("op_5123_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5131_promoted_to_fp16 = const()[name = string("op_5131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_5123_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5133_cast_fp16 = pow(x = x_149_cast_fp16, y = var_5131_promoted_to_fp16)[name = string("op_5133_cast_fp16")];
            tensor<int32, [1]> var_5135_axes_0 = const()[name = string("op_5135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5135_keep_dims_0 = const()[name = string("op_5135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5135_cast_fp16 = reduce_mean(axes = var_5135_axes_0, keep_dims = var_5135_keep_dims_0, x = var_5133_cast_fp16)[name = string("op_5135_cast_fp16")];
            fp16 var_5136_to_fp16 = const()[name = string("op_5136_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5137_cast_fp16 = add(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5138_cast_fp16 = rsqrt(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5139_cast_fp16 = mul(x = x_149_cast_fp16, y = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503510912)))];
            tensor<fp16, [1, 8, 16, 128]> x_153_cast_fp16 = mul(x = var_5139_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503511232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504559872))))[name = string("model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5148 = const()[name = string("op_5148"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5149_cast_fp16 = reshape(shape = var_5148, x = linear_86_cast_fp16)[name = string("op_5149_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_49_begin_0 = const()[name = string("first_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_49_end_0 = const()[name = string("first_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_49_end_mask_0 = const()[name = string("first_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_49_cast_fp16 = slice_by_index(begin = first_49_begin_0, end = first_49_end_0, end_mask = first_49_end_mask_0, x = x_151_cast_fp16)[name = string("first_49_cast_fp16")];
            tensor<int32, [4]> second_49_begin_0 = const()[name = string("second_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_49_end_0 = const()[name = string("second_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_49_end_mask_0 = const()[name = string("second_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_49_cast_fp16 = slice_by_index(begin = second_49_begin_0, end = second_49_end_0, end_mask = second_49_end_mask_0, x = x_151_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5163_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5164_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5164_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5166_cast_fp16 = sub(x = var_5163_cast_fp16, y = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5167_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5168_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5170_cast_fp16 = add(x = var_5167_cast_fp16, y = var_5168_cast_fp16)[name = string("op_5170_cast_fp16")];
            int32 var_5172 = const()[name = string("op_5172"), val = int32(-1)];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_25_cast_fp16 = concat(axis = var_5172, interleave = q_25_interleave_0, values = (var_5166_cast_fp16, var_5170_cast_fp16))[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> first_51_begin_0 = const()[name = string("first_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_51_end_0 = const()[name = string("first_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_51_end_mask_0 = const()[name = string("first_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_51_cast_fp16 = slice_by_index(begin = first_51_begin_0, end = first_51_end_0, end_mask = first_51_end_mask_0, x = x_153_cast_fp16)[name = string("first_51_cast_fp16")];
            tensor<int32, [4]> second_51_begin_0 = const()[name = string("second_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_51_end_0 = const()[name = string("second_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_51_end_mask_0 = const()[name = string("second_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_51_cast_fp16 = slice_by_index(begin = second_51_begin_0, end = second_51_end_0, end_mask = second_51_end_mask_0, x = x_153_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5184_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5185_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5185_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5187_cast_fp16 = sub(x = var_5184_cast_fp16, y = var_5185_cast_fp16)[name = string("op_5187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5188_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5189_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5191_cast_fp16 = add(x = var_5188_cast_fp16, y = var_5189_cast_fp16)[name = string("op_5191_cast_fp16")];
            int32 var_5193 = const()[name = string("op_5193"), val = int32(-1)];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_25_cast_fp16 = concat(axis = var_5193, interleave = k_25_interleave_0, values = (var_5187_cast_fp16, var_5191_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5204_begin_0 = const()[name = string("op_5204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5204_end_0 = const()[name = string("op_5204_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5204_end_mask_0 = const()[name = string("op_5204_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5204_stride_0 = const()[name = string("op_5204_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5204_cast_fp16 = slice_by_index(begin = var_5204_begin_0, end = var_5204_end_0, end_mask = var_5204_end_mask_0, stride = var_5204_stride_0, x = q_25_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5214_stride_0 = const()[name = string("op_5214_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, stride = var_5214_stride_0, x = q_25_cast_fp16)[name = string("op_5214_cast_fp16")];
            int32 var_5216 = const()[name = string("op_5216"), val = int32(2)];
            bool stacked_25_interleave_0 = const()[name = string("stacked_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_25_cast_fp16 = concat(axis = var_5216, interleave = stacked_25_interleave_0, values = (var_5204_cast_fp16, var_5214_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5222_begin_0 = const()[name = string("op_5222_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5222_end_0 = const()[name = string("op_5222_end_0"), val = tensor<int32, [4]>([13, 8, 2064, 128])];
            tensor<bool, [4]> var_5222_end_mask_0 = const()[name = string("op_5222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, x = read_state_0)[name = string("op_5222_cast_fp16")];
            tensor<int32, [4]> var_5232_begin_0 = const()[name = string("op_5232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5232_end_0 = const()[name = string("op_5232_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5232_end_mask_0 = const()[name = string("op_5232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5232_cast_fp16 = slice_by_index(begin = var_5232_begin_0, end = var_5232_end_0, end_mask = var_5232_end_mask_0, x = var_5222_cast_fp16)[name = string("op_5232_cast_fp16")];
            bool var_5241_transpose_x_1 = const()[name = string("op_5241_transpose_x_1"), val = bool(false)];
            bool var_5241_transpose_y_1 = const()[name = string("op_5241_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5241_cast_fp16 = matmul(transpose_x = var_5241_transpose_x_1, transpose_y = var_5241_transpose_y_1, x = var_5232_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5241_cast_fp16")];
            tensor<int32, [4]> var_5244_perm_0 = const()[name = string("op_5244_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5245_to_fp16 = const()[name = string("op_5245_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5244_cast_fp16 = transpose(perm = var_5244_perm_0, x = var_5241_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 2048]> var_5246_cast_fp16 = mul(x = var_5244_cast_fp16, y = var_5245_to_fp16)[name = string("op_5246_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_25_cast_fp16 = add(x = var_5246_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5252_transpose_x_1 = const()[name = string("op_5252_transpose_x_1"), val = bool(false)];
            bool var_5252_transpose_y_1 = const()[name = string("op_5252_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5252_cast_fp16 = matmul(transpose_x = var_5252_transpose_x_1, transpose_y = var_5252_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5252_cast_fp16")];
            fp16 var_5253_to_fp16 = const()[name = string("op_5253_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5254_cast_fp16 = mul(x = var_5252_cast_fp16, y = var_5253_to_fp16)[name = string("op_5254_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5254_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5258 = const()[name = string("op_5258"), val = int32(-1)];
            bool var_5259_interleave_0 = const()[name = string("op_5259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5259_cast_fp16 = concat(axis = var_5258, interleave = var_5259_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5259_cast_fp16")];
            int32 var_5260 = const()[name = string("op_5260"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_25_cast_fp16 = softmax(axis = var_5260, x = var_5259_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5267_begin_0 = const()[name = string("op_5267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5267_end_0 = const()[name = string("op_5267_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5267_end_mask_0 = const()[name = string("op_5267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5267_cast_fp16 = slice_by_index(begin = var_5267_begin_0, end = var_5267_end_0, end_mask = var_5267_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<int32, [4]> var_5272_begin_0 = const()[name = string("op_5272_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5272_end_0 = const()[name = string("op_5272_end_0"), val = tensor<int32, [4]>([13, 8, 2064, 128])];
            tensor<bool, [4]> var_5272_end_mask_0 = const()[name = string("op_5272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5272_cast_fp16 = slice_by_index(begin = var_5272_begin_0, end = var_5272_end_0, end_mask = var_5272_end_mask_0, x = read_state_1)[name = string("op_5272_cast_fp16")];
            tensor<int32, [4]> var_5282_begin_0 = const()[name = string("op_5282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5282_end_0 = const()[name = string("op_5282_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5282_end_mask_0 = const()[name = string("op_5282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5282_cast_fp16 = slice_by_index(begin = var_5282_begin_0, end = var_5282_end_0, end_mask = var_5282_end_mask_0, x = var_5272_cast_fp16)[name = string("op_5282_cast_fp16")];
            bool var_5288_transpose_x_0 = const()[name = string("op_5288_transpose_x_0"), val = bool(false)];
            bool var_5288_transpose_y_0 = const()[name = string("op_5288_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5288_cast_fp16 = matmul(transpose_x = var_5288_transpose_x_0, transpose_y = var_5288_transpose_y_0, x = var_5267_cast_fp16, y = var_5282_cast_fp16)[name = string("op_5288_cast_fp16")];
            tensor<int32, [4]> var_5293_begin_0 = const()[name = string("op_5293_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5293_end_0 = const()[name = string("op_5293_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5293_end_mask_0 = const()[name = string("op_5293_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5293_cast_fp16 = slice_by_index(begin = var_5293_begin_0, end = var_5293_end_0, end_mask = var_5293_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5293_cast_fp16")];
            bool var_5294_transpose_x_0 = const()[name = string("op_5294_transpose_x_0"), val = bool(false)];
            bool var_5294_transpose_y_0 = const()[name = string("op_5294_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5149_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5294_cast_fp16 = matmul(transpose_x = var_5294_transpose_x_0, transpose_y = var_5294_transpose_y_0, x = var_5293_cast_fp16, y = v_25_cast_fp16)[name = string("op_5294_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5288_cast_fp16, y = var_5294_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5311_begin_0 = const()[name = string("op_5311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5311_end_0 = const()[name = string("op_5311_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5311_end_mask_0 = const()[name = string("op_5311_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5311_cast_fp16 = slice_by_index(begin = var_5311_begin_0, end = var_5311_end_0, end_mask = var_5311_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> var_5326_begin_0 = const()[name = string("op_5326_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5326_end_0 = const()[name = string("op_5326_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5326_end_mask_0 = const()[name = string("op_5326_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5326_cast_fp16 = slice_by_index(begin = var_5326_begin_0, end = var_5326_end_0, end_mask = var_5326_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5326_cast_fp16")];
            int32 var_5328 = const()[name = string("op_5328"), val = int32(1)];
            bool mixed_51_interleave_0 = const()[name = string("mixed_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_51_cast_fp16 = concat(axis = var_5328, interleave = mixed_51_interleave_0, values = (var_5311_cast_fp16, var_5326_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504561984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506659200))))[name = string("model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = linear_87_cast_fp16)[name = string("x_155_cast_fp16")];
            fp16 var_5347_promoted_to_fp16 = const()[name = string("op_5347_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5349_cast_fp16 = pow(x = x_155_cast_fp16, y = var_5347_promoted_to_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5351_axes_0 = const()[name = string("op_5351_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5351_keep_dims_0 = const()[name = string("op_5351_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5351_cast_fp16 = reduce_mean(axes = var_5351_axes_0, keep_dims = var_5351_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5351_cast_fp16")];
            fp16 var_5352_to_fp16 = const()[name = string("op_5352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5353_cast_fp16 = add(x = var_5351_cast_fp16, y = var_5352_to_fp16)[name = string("op_5353_cast_fp16")];
            fp32 var_5354_epsilon_0 = const()[name = string("op_5354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5354_cast_fp16 = rsqrt(epsilon = var_5354_epsilon_0, x = var_5353_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5355_cast_fp16 = mul(x = x_155_cast_fp16, y = var_5354_cast_fp16)[name = string("op_5355_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506661312)))];
            tensor<fp16, [1, 16, 1024]> input_125_cast_fp16 = mul(x = var_5355_cast_fp16, y = model_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506663424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509809216))))[name = string("model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5360_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509815424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512961216))))[name = string("model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_129_cast_fp16 = mul(x = var_5360_cast_fp16, y = linear_89_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512967424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516113216))))[name = string("model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_157_cast_fp16 = add(x = x_155_cast_fp16, y = linear_90_cast_fp16)[name = string("x_157_cast_fp16")];
            fp16 var_5374_promoted_to_fp16 = const()[name = string("op_5374_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5376_cast_fp16 = pow(x = x_157_cast_fp16, y = var_5374_promoted_to_fp16)[name = string("op_5376_cast_fp16")];
            tensor<int32, [1]> var_5378_axes_0 = const()[name = string("op_5378_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5378_keep_dims_0 = const()[name = string("op_5378_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5378_cast_fp16 = reduce_mean(axes = var_5378_axes_0, keep_dims = var_5378_keep_dims_0, x = var_5376_cast_fp16)[name = string("op_5378_cast_fp16")];
            fp16 var_5379_to_fp16 = const()[name = string("op_5379_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5380_cast_fp16 = add(x = var_5378_cast_fp16, y = var_5379_to_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5381_cast_fp16 = rsqrt(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5382_cast_fp16 = mul(x = x_157_cast_fp16, y = var_5381_cast_fp16)[name = string("op_5382_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516115328)))];
            tensor<fp16, [1, 16, 1024]> input_131_cast_fp16 = mul(x = var_5382_cast_fp16, y = model_layers_13_input_layernorm_weight_to_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516117440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518214656))))[name = string("model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5391 = const()[name = string("op_5391"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5392_cast_fp16 = reshape(shape = var_5391, x = linear_91_cast_fp16)[name = string("op_5392_cast_fp16")];
            tensor<int32, [4]> x_159_perm_0 = const()[name = string("x_159_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5400_promoted_to_fp16 = const()[name = string("op_5400_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_159_cast_fp16 = transpose(perm = x_159_perm_0, x = var_5392_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5402_cast_fp16 = pow(x = x_159_cast_fp16, y = var_5400_promoted_to_fp16)[name = string("op_5402_cast_fp16")];
            tensor<int32, [1]> var_5404_axes_0 = const()[name = string("op_5404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5404_keep_dims_0 = const()[name = string("op_5404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5404_cast_fp16 = reduce_mean(axes = var_5404_axes_0, keep_dims = var_5404_keep_dims_0, x = var_5402_cast_fp16)[name = string("op_5404_cast_fp16")];
            fp16 var_5405_to_fp16 = const()[name = string("op_5405_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5406_cast_fp16 = add(x = var_5404_cast_fp16, y = var_5405_to_fp16)[name = string("op_5406_cast_fp16")];
            fp32 var_5407_epsilon_0 = const()[name = string("op_5407_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5407_cast_fp16 = rsqrt(epsilon = var_5407_epsilon_0, x = var_5406_cast_fp16)[name = string("op_5407_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5408_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5407_cast_fp16)[name = string("op_5408_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518218816)))];
            tensor<fp16, [1, 16, 16, 128]> x_163_cast_fp16 = mul(x = var_5408_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518219136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519267776))))[name = string("model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5417 = const()[name = string("op_5417"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5418_cast_fp16 = reshape(shape = var_5417, x = linear_92_cast_fp16)[name = string("op_5418_cast_fp16")];
            tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5426_promoted_to_fp16 = const()[name = string("op_5426_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_5418_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5428_cast_fp16 = pow(x = x_161_cast_fp16, y = var_5426_promoted_to_fp16)[name = string("op_5428_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5428_cast_fp16)[name = string("op_5430_cast_fp16")];
            fp16 var_5431_to_fp16 = const()[name = string("op_5431_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5432_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_to_fp16)[name = string("op_5432_cast_fp16")];
            fp32 var_5433_epsilon_0 = const()[name = string("op_5433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5433_cast_fp16 = rsqrt(epsilon = var_5433_epsilon_0, x = var_5432_cast_fp16)[name = string("op_5433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5434_cast_fp16 = mul(x = x_161_cast_fp16, y = var_5433_cast_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519269888)))];
            tensor<fp16, [1, 8, 16, 128]> x_165_cast_fp16 = mul(x = var_5434_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519270208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520318848))))[name = string("model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5443 = const()[name = string("op_5443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5444_cast_fp16 = reshape(shape = var_5443, x = linear_93_cast_fp16)[name = string("op_5444_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_53_begin_0 = const()[name = string("first_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_53_end_0 = const()[name = string("first_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_53_end_mask_0 = const()[name = string("first_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_53_cast_fp16 = slice_by_index(begin = first_53_begin_0, end = first_53_end_0, end_mask = first_53_end_mask_0, x = x_163_cast_fp16)[name = string("first_53_cast_fp16")];
            tensor<int32, [4]> second_53_begin_0 = const()[name = string("second_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_53_end_0 = const()[name = string("second_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_53_end_mask_0 = const()[name = string("second_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_53_cast_fp16 = slice_by_index(begin = second_53_begin_0, end = second_53_end_0, end_mask = second_53_end_mask_0, x = x_163_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5458_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5459_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5461_cast_fp16 = sub(x = var_5458_cast_fp16, y = var_5459_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5462_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5463_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5465_cast_fp16 = add(x = var_5462_cast_fp16, y = var_5463_cast_fp16)[name = string("op_5465_cast_fp16")];
            int32 var_5467 = const()[name = string("op_5467"), val = int32(-1)];
            bool q_27_interleave_0 = const()[name = string("q_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_27_cast_fp16 = concat(axis = var_5467, interleave = q_27_interleave_0, values = (var_5461_cast_fp16, var_5465_cast_fp16))[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> first_55_begin_0 = const()[name = string("first_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_55_end_0 = const()[name = string("first_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_55_end_mask_0 = const()[name = string("first_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_55_cast_fp16 = slice_by_index(begin = first_55_begin_0, end = first_55_end_0, end_mask = first_55_end_mask_0, x = x_165_cast_fp16)[name = string("first_55_cast_fp16")];
            tensor<int32, [4]> second_55_begin_0 = const()[name = string("second_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_55_end_0 = const()[name = string("second_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_55_end_mask_0 = const()[name = string("second_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_55_cast_fp16 = slice_by_index(begin = second_55_begin_0, end = second_55_end_0, end_mask = second_55_end_mask_0, x = x_165_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5479_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5480_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5482_cast_fp16 = sub(x = var_5479_cast_fp16, y = var_5480_cast_fp16)[name = string("op_5482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5483_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5484_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5486_cast_fp16 = add(x = var_5483_cast_fp16, y = var_5484_cast_fp16)[name = string("op_5486_cast_fp16")];
            int32 var_5488 = const()[name = string("op_5488"), val = int32(-1)];
            bool k_27_interleave_0 = const()[name = string("k_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_27_cast_fp16 = concat(axis = var_5488, interleave = k_27_interleave_0, values = (var_5482_cast_fp16, var_5486_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5499_begin_0 = const()[name = string("op_5499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5499_end_0 = const()[name = string("op_5499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5499_end_mask_0 = const()[name = string("op_5499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5499_stride_0 = const()[name = string("op_5499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5499_cast_fp16 = slice_by_index(begin = var_5499_begin_0, end = var_5499_end_0, end_mask = var_5499_end_mask_0, stride = var_5499_stride_0, x = q_27_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<int32, [4]> var_5509_begin_0 = const()[name = string("op_5509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5509_end_0 = const()[name = string("op_5509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5509_end_mask_0 = const()[name = string("op_5509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5509_stride_0 = const()[name = string("op_5509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5509_cast_fp16 = slice_by_index(begin = var_5509_begin_0, end = var_5509_end_0, end_mask = var_5509_end_mask_0, stride = var_5509_stride_0, x = q_27_cast_fp16)[name = string("op_5509_cast_fp16")];
            int32 var_5511 = const()[name = string("op_5511"), val = int32(2)];
            bool stacked_27_interleave_0 = const()[name = string("stacked_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_27_cast_fp16 = concat(axis = var_5511, interleave = stacked_27_interleave_0, values = (var_5499_cast_fp16, var_5509_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5517_begin_0 = const()[name = string("op_5517_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5517_end_0 = const()[name = string("op_5517_end_0"), val = tensor<int32, [4]>([14, 8, 2064, 128])];
            tensor<bool, [4]> var_5517_end_mask_0 = const()[name = string("op_5517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5517_cast_fp16 = slice_by_index(begin = var_5517_begin_0, end = var_5517_end_0, end_mask = var_5517_end_mask_0, x = read_state_0)[name = string("op_5517_cast_fp16")];
            tensor<int32, [4]> var_5527_begin_0 = const()[name = string("op_5527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5527_end_0 = const()[name = string("op_5527_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5527_end_mask_0 = const()[name = string("op_5527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5527_cast_fp16 = slice_by_index(begin = var_5527_begin_0, end = var_5527_end_0, end_mask = var_5527_end_mask_0, x = var_5517_cast_fp16)[name = string("op_5527_cast_fp16")];
            bool var_5536_transpose_x_1 = const()[name = string("op_5536_transpose_x_1"), val = bool(false)];
            bool var_5536_transpose_y_1 = const()[name = string("op_5536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5536_cast_fp16 = matmul(transpose_x = var_5536_transpose_x_1, transpose_y = var_5536_transpose_y_1, x = var_5527_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5536_cast_fp16")];
            tensor<int32, [4]> var_5539_perm_0 = const()[name = string("op_5539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5540_to_fp16 = const()[name = string("op_5540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5539_cast_fp16 = transpose(perm = var_5539_perm_0, x = var_5536_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 2048]> var_5541_cast_fp16 = mul(x = var_5539_cast_fp16, y = var_5540_to_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_27_cast_fp16 = add(x = var_5541_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5547_transpose_x_1 = const()[name = string("op_5547_transpose_x_1"), val = bool(false)];
            bool var_5547_transpose_y_1 = const()[name = string("op_5547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5547_cast_fp16 = matmul(transpose_x = var_5547_transpose_x_1, transpose_y = var_5547_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5547_cast_fp16")];
            fp16 var_5548_to_fp16 = const()[name = string("op_5548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5549_cast_fp16 = mul(x = var_5547_cast_fp16, y = var_5548_to_fp16)[name = string("op_5549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5553 = const()[name = string("op_5553"), val = int32(-1)];
            bool var_5554_interleave_0 = const()[name = string("op_5554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5554_cast_fp16 = concat(axis = var_5553, interleave = var_5554_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5554_cast_fp16")];
            int32 var_5555 = const()[name = string("op_5555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_27_cast_fp16 = softmax(axis = var_5555, x = var_5554_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5562_begin_0 = const()[name = string("op_5562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5562_end_0 = const()[name = string("op_5562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5562_end_mask_0 = const()[name = string("op_5562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5562_cast_fp16 = slice_by_index(begin = var_5562_begin_0, end = var_5562_end_0, end_mask = var_5562_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5562_cast_fp16")];
            tensor<int32, [4]> var_5567_begin_0 = const()[name = string("op_5567_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5567_end_0 = const()[name = string("op_5567_end_0"), val = tensor<int32, [4]>([14, 8, 2064, 128])];
            tensor<bool, [4]> var_5567_end_mask_0 = const()[name = string("op_5567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5567_cast_fp16 = slice_by_index(begin = var_5567_begin_0, end = var_5567_end_0, end_mask = var_5567_end_mask_0, x = read_state_1)[name = string("op_5567_cast_fp16")];
            tensor<int32, [4]> var_5577_begin_0 = const()[name = string("op_5577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5577_end_0 = const()[name = string("op_5577_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5577_end_mask_0 = const()[name = string("op_5577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5577_cast_fp16 = slice_by_index(begin = var_5577_begin_0, end = var_5577_end_0, end_mask = var_5577_end_mask_0, x = var_5567_cast_fp16)[name = string("op_5577_cast_fp16")];
            bool var_5583_transpose_x_0 = const()[name = string("op_5583_transpose_x_0"), val = bool(false)];
            bool var_5583_transpose_y_0 = const()[name = string("op_5583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5583_cast_fp16 = matmul(transpose_x = var_5583_transpose_x_0, transpose_y = var_5583_transpose_y_0, x = var_5562_cast_fp16, y = var_5577_cast_fp16)[name = string("op_5583_cast_fp16")];
            tensor<int32, [4]> var_5588_begin_0 = const()[name = string("op_5588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5588_end_0 = const()[name = string("op_5588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5588_end_mask_0 = const()[name = string("op_5588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5588_cast_fp16 = slice_by_index(begin = var_5588_begin_0, end = var_5588_end_0, end_mask = var_5588_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5588_cast_fp16")];
            bool var_5589_transpose_x_0 = const()[name = string("op_5589_transpose_x_0"), val = bool(false)];
            bool var_5589_transpose_y_0 = const()[name = string("op_5589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5444_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5589_cast_fp16 = matmul(transpose_x = var_5589_transpose_x_0, transpose_y = var_5589_transpose_y_0, x = var_5588_cast_fp16, y = v_27_cast_fp16)[name = string("op_5589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5583_cast_fp16, y = var_5589_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5606_begin_0 = const()[name = string("op_5606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5606_end_0 = const()[name = string("op_5606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5606_end_mask_0 = const()[name = string("op_5606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5606_cast_fp16 = slice_by_index(begin = var_5606_begin_0, end = var_5606_end_0, end_mask = var_5606_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5621_begin_0 = const()[name = string("op_5621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5621_end_0 = const()[name = string("op_5621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5621_end_mask_0 = const()[name = string("op_5621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5621_cast_fp16 = slice_by_index(begin = var_5621_begin_0, end = var_5621_end_0, end_mask = var_5621_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5621_cast_fp16")];
            int32 var_5623 = const()[name = string("op_5623"), val = int32(1)];
            bool mixed_55_interleave_0 = const()[name = string("mixed_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_55_cast_fp16 = concat(axis = var_5623, interleave = mixed_55_interleave_0, values = (var_5606_cast_fp16, var_5621_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5631 = const()[name = string("op_5631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_5631, x = var_5627_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520320960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522418176))))[name = string("model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_167_cast_fp16 = add(x = x_157_cast_fp16, y = linear_94_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_5642_promoted_to_fp16 = const()[name = string("op_5642_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5644_cast_fp16 = pow(x = x_167_cast_fp16, y = var_5642_promoted_to_fp16)[name = string("op_5644_cast_fp16")];
            tensor<int32, [1]> var_5646_axes_0 = const()[name = string("op_5646_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5646_keep_dims_0 = const()[name = string("op_5646_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5646_cast_fp16 = reduce_mean(axes = var_5646_axes_0, keep_dims = var_5646_keep_dims_0, x = var_5644_cast_fp16)[name = string("op_5646_cast_fp16")];
            fp16 var_5647_to_fp16 = const()[name = string("op_5647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5648_cast_fp16 = add(x = var_5646_cast_fp16, y = var_5647_to_fp16)[name = string("op_5648_cast_fp16")];
            fp32 var_5649_epsilon_0 = const()[name = string("op_5649_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5649_cast_fp16 = rsqrt(epsilon = var_5649_epsilon_0, x = var_5648_cast_fp16)[name = string("op_5649_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5650_cast_fp16 = mul(x = x_167_cast_fp16, y = var_5649_cast_fp16)[name = string("op_5650_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522420288)))];
            tensor<fp16, [1, 16, 1024]> input_135_cast_fp16 = mul(x = var_5650_cast_fp16, y = model_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522422400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525568192))))[name = string("model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5655_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_5655_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525574400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528720192))))[name = string("model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_139_cast_fp16 = mul(x = var_5655_cast_fp16, y = linear_96_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528726400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531872192))))[name = string("model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_169_cast_fp16 = add(x = x_167_cast_fp16, y = linear_97_cast_fp16)[name = string("x_169_cast_fp16")];
            fp16 var_5669_promoted_to_fp16 = const()[name = string("op_5669_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5671_cast_fp16 = pow(x = x_169_cast_fp16, y = var_5669_promoted_to_fp16)[name = string("op_5671_cast_fp16")];
            tensor<int32, [1]> var_5673_axes_0 = const()[name = string("op_5673_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5673_keep_dims_0 = const()[name = string("op_5673_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5673_cast_fp16 = reduce_mean(axes = var_5673_axes_0, keep_dims = var_5673_keep_dims_0, x = var_5671_cast_fp16)[name = string("op_5673_cast_fp16")];
            fp16 var_5674_to_fp16 = const()[name = string("op_5674_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5675_cast_fp16 = add(x = var_5673_cast_fp16, y = var_5674_to_fp16)[name = string("op_5675_cast_fp16")];
            fp32 var_5676_epsilon_0 = const()[name = string("op_5676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5676_cast_fp16 = rsqrt(epsilon = var_5676_epsilon_0, x = var_5675_cast_fp16)[name = string("op_5676_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5677_cast_fp16 = mul(x = x_169_cast_fp16, y = var_5676_cast_fp16)[name = string("op_5677_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531874304)))];
            tensor<fp16, [1, 16, 1024]> input_141_cast_fp16 = mul(x = var_5677_cast_fp16, y = model_layers_14_input_layernorm_weight_to_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531876416))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533973632))))[name = string("model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_5686 = const()[name = string("op_5686"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5687_cast_fp16 = reshape(shape = var_5686, x = linear_98_cast_fp16)[name = string("op_5687_cast_fp16")];
            tensor<int32, [4]> x_171_perm_0 = const()[name = string("x_171_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5695_promoted_to_fp16 = const()[name = string("op_5695_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_171_cast_fp16 = transpose(perm = x_171_perm_0, x = var_5687_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_5697_cast_fp16 = pow(x = x_171_cast_fp16, y = var_5695_promoted_to_fp16)[name = string("op_5697_cast_fp16")];
            tensor<int32, [1]> var_5699_axes_0 = const()[name = string("op_5699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5699_keep_dims_0 = const()[name = string("op_5699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5699_cast_fp16 = reduce_mean(axes = var_5699_axes_0, keep_dims = var_5699_keep_dims_0, x = var_5697_cast_fp16)[name = string("op_5699_cast_fp16")];
            fp16 var_5700_to_fp16 = const()[name = string("op_5700_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5701_cast_fp16 = add(x = var_5699_cast_fp16, y = var_5700_to_fp16)[name = string("op_5701_cast_fp16")];
            fp32 var_5702_epsilon_0 = const()[name = string("op_5702_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5702_cast_fp16 = rsqrt(epsilon = var_5702_epsilon_0, x = var_5701_cast_fp16)[name = string("op_5702_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5703_cast_fp16 = mul(x = x_171_cast_fp16, y = var_5702_cast_fp16)[name = string("op_5703_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533977792)))];
            tensor<fp16, [1, 16, 16, 128]> x_175_cast_fp16 = mul(x = var_5703_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_175_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533978112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535026752))))[name = string("model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_5712 = const()[name = string("op_5712"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5713_cast_fp16 = reshape(shape = var_5712, x = linear_99_cast_fp16)[name = string("op_5713_cast_fp16")];
            tensor<int32, [4]> x_173_perm_0 = const()[name = string("x_173_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5721_promoted_to_fp16 = const()[name = string("op_5721_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_173_cast_fp16 = transpose(perm = x_173_perm_0, x = var_5713_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_5723_cast_fp16 = pow(x = x_173_cast_fp16, y = var_5721_promoted_to_fp16)[name = string("op_5723_cast_fp16")];
            tensor<int32, [1]> var_5725_axes_0 = const()[name = string("op_5725_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5725_keep_dims_0 = const()[name = string("op_5725_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5725_cast_fp16 = reduce_mean(axes = var_5725_axes_0, keep_dims = var_5725_keep_dims_0, x = var_5723_cast_fp16)[name = string("op_5725_cast_fp16")];
            fp16 var_5726_to_fp16 = const()[name = string("op_5726_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5727_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5726_to_fp16)[name = string("op_5727_cast_fp16")];
            fp32 var_5728_epsilon_0 = const()[name = string("op_5728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5728_cast_fp16 = rsqrt(epsilon = var_5728_epsilon_0, x = var_5727_cast_fp16)[name = string("op_5728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5729_cast_fp16 = mul(x = x_173_cast_fp16, y = var_5728_cast_fp16)[name = string("op_5729_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535028864)))];
            tensor<fp16, [1, 8, 16, 128]> x_177_cast_fp16 = mul(x = var_5729_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535029184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536077824))))[name = string("model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5738 = const()[name = string("op_5738"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5739_cast_fp16 = reshape(shape = var_5738, x = linear_100_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_57_begin_0 = const()[name = string("first_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_57_end_0 = const()[name = string("first_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_57_end_mask_0 = const()[name = string("first_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_57_cast_fp16 = slice_by_index(begin = first_57_begin_0, end = first_57_end_0, end_mask = first_57_end_mask_0, x = x_175_cast_fp16)[name = string("first_57_cast_fp16")];
            tensor<int32, [4]> second_57_begin_0 = const()[name = string("second_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_57_end_0 = const()[name = string("second_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_57_end_mask_0 = const()[name = string("second_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_57_cast_fp16 = slice_by_index(begin = second_57_begin_0, end = second_57_end_0, end_mask = second_57_end_mask_0, x = x_175_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5753_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5754_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5756_cast_fp16 = sub(x = var_5753_cast_fp16, y = var_5754_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5757_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5757_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5758_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5758_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5760_cast_fp16 = add(x = var_5757_cast_fp16, y = var_5758_cast_fp16)[name = string("op_5760_cast_fp16")];
            int32 var_5762 = const()[name = string("op_5762"), val = int32(-1)];
            bool q_29_interleave_0 = const()[name = string("q_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_29_cast_fp16 = concat(axis = var_5762, interleave = q_29_interleave_0, values = (var_5756_cast_fp16, var_5760_cast_fp16))[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> first_59_begin_0 = const()[name = string("first_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_59_end_0 = const()[name = string("first_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_59_end_mask_0 = const()[name = string("first_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_59_cast_fp16 = slice_by_index(begin = first_59_begin_0, end = first_59_end_0, end_mask = first_59_end_mask_0, x = x_177_cast_fp16)[name = string("first_59_cast_fp16")];
            tensor<int32, [4]> second_59_begin_0 = const()[name = string("second_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_59_end_0 = const()[name = string("second_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_59_end_mask_0 = const()[name = string("second_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_59_cast_fp16 = slice_by_index(begin = second_59_begin_0, end = second_59_end_0, end_mask = second_59_end_mask_0, x = x_177_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5774_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5775_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5777_cast_fp16 = sub(x = var_5774_cast_fp16, y = var_5775_cast_fp16)[name = string("op_5777_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5778_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5779_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5779_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5781_cast_fp16 = add(x = var_5778_cast_fp16, y = var_5779_cast_fp16)[name = string("op_5781_cast_fp16")];
            int32 var_5783 = const()[name = string("op_5783"), val = int32(-1)];
            bool k_29_interleave_0 = const()[name = string("k_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_29_cast_fp16 = concat(axis = var_5783, interleave = k_29_interleave_0, values = (var_5777_cast_fp16, var_5781_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_5794_begin_0 = const()[name = string("op_5794_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5794_end_0 = const()[name = string("op_5794_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5794_end_mask_0 = const()[name = string("op_5794_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5794_stride_0 = const()[name = string("op_5794_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5794_cast_fp16 = slice_by_index(begin = var_5794_begin_0, end = var_5794_end_0, end_mask = var_5794_end_mask_0, stride = var_5794_stride_0, x = q_29_cast_fp16)[name = string("op_5794_cast_fp16")];
            tensor<int32, [4]> var_5804_begin_0 = const()[name = string("op_5804_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5804_end_0 = const()[name = string("op_5804_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5804_end_mask_0 = const()[name = string("op_5804_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5804_stride_0 = const()[name = string("op_5804_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5804_cast_fp16 = slice_by_index(begin = var_5804_begin_0, end = var_5804_end_0, end_mask = var_5804_end_mask_0, stride = var_5804_stride_0, x = q_29_cast_fp16)[name = string("op_5804_cast_fp16")];
            int32 var_5806 = const()[name = string("op_5806"), val = int32(2)];
            bool stacked_29_interleave_0 = const()[name = string("stacked_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_29_cast_fp16 = concat(axis = var_5806, interleave = stacked_29_interleave_0, values = (var_5794_cast_fp16, var_5804_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_5812_begin_0 = const()[name = string("op_5812_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5812_end_0 = const()[name = string("op_5812_end_0"), val = tensor<int32, [4]>([15, 8, 2064, 128])];
            tensor<bool, [4]> var_5812_end_mask_0 = const()[name = string("op_5812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5812_cast_fp16 = slice_by_index(begin = var_5812_begin_0, end = var_5812_end_0, end_mask = var_5812_end_mask_0, x = read_state_0)[name = string("op_5812_cast_fp16")];
            tensor<int32, [4]> var_5822_begin_0 = const()[name = string("op_5822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5822_end_0 = const()[name = string("op_5822_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5822_end_mask_0 = const()[name = string("op_5822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5822_cast_fp16 = slice_by_index(begin = var_5822_begin_0, end = var_5822_end_0, end_mask = var_5822_end_mask_0, x = var_5812_cast_fp16)[name = string("op_5822_cast_fp16")];
            bool var_5831_transpose_x_1 = const()[name = string("op_5831_transpose_x_1"), val = bool(false)];
            bool var_5831_transpose_y_1 = const()[name = string("op_5831_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_5831_cast_fp16 = matmul(transpose_x = var_5831_transpose_x_1, transpose_y = var_5831_transpose_y_1, x = var_5822_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_5831_cast_fp16")];
            tensor<int32, [4]> var_5834_perm_0 = const()[name = string("op_5834_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5835_to_fp16 = const()[name = string("op_5835_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_5834_cast_fp16 = transpose(perm = var_5834_perm_0, x = var_5831_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 2048]> var_5836_cast_fp16 = mul(x = var_5834_cast_fp16, y = var_5835_to_fp16)[name = string("op_5836_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_29_cast_fp16 = add(x = var_5836_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_5842_transpose_x_1 = const()[name = string("op_5842_transpose_x_1"), val = bool(false)];
            bool var_5842_transpose_y_1 = const()[name = string("op_5842_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5842_cast_fp16 = matmul(transpose_x = var_5842_transpose_x_1, transpose_y = var_5842_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_5842_cast_fp16")];
            fp16 var_5843_to_fp16 = const()[name = string("op_5843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5844_cast_fp16 = mul(x = var_5842_cast_fp16, y = var_5843_to_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_5844_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_5848 = const()[name = string("op_5848"), val = int32(-1)];
            bool var_5849_interleave_0 = const()[name = string("op_5849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_5849_cast_fp16 = concat(axis = var_5848, interleave = var_5849_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_5849_cast_fp16")];
            int32 var_5850 = const()[name = string("op_5850"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_29_cast_fp16 = softmax(axis = var_5850, x = var_5849_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_5857_begin_0 = const()[name = string("op_5857_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5857_end_0 = const()[name = string("op_5857_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_5857_end_mask_0 = const()[name = string("op_5857_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_5857_cast_fp16 = slice_by_index(begin = var_5857_begin_0, end = var_5857_end_0, end_mask = var_5857_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5857_cast_fp16")];
            tensor<int32, [4]> var_5862_begin_0 = const()[name = string("op_5862_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5862_end_0 = const()[name = string("op_5862_end_0"), val = tensor<int32, [4]>([15, 8, 2064, 128])];
            tensor<bool, [4]> var_5862_end_mask_0 = const()[name = string("op_5862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_5862_cast_fp16 = slice_by_index(begin = var_5862_begin_0, end = var_5862_end_0, end_mask = var_5862_end_mask_0, x = read_state_1)[name = string("op_5862_cast_fp16")];
            tensor<int32, [4]> var_5872_begin_0 = const()[name = string("op_5872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5872_end_0 = const()[name = string("op_5872_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_5872_end_mask_0 = const()[name = string("op_5872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_5872_cast_fp16 = slice_by_index(begin = var_5872_begin_0, end = var_5872_end_0, end_mask = var_5872_end_mask_0, x = var_5862_cast_fp16)[name = string("op_5872_cast_fp16")];
            bool var_5878_transpose_x_0 = const()[name = string("op_5878_transpose_x_0"), val = bool(false)];
            bool var_5878_transpose_y_0 = const()[name = string("op_5878_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5878_cast_fp16 = matmul(transpose_x = var_5878_transpose_x_0, transpose_y = var_5878_transpose_y_0, x = var_5857_cast_fp16, y = var_5872_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<int32, [4]> var_5883_begin_0 = const()[name = string("op_5883_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_5883_end_0 = const()[name = string("op_5883_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_5883_end_mask_0 = const()[name = string("op_5883_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5883_cast_fp16 = slice_by_index(begin = var_5883_begin_0, end = var_5883_end_0, end_mask = var_5883_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5883_cast_fp16")];
            bool var_5884_transpose_x_0 = const()[name = string("op_5884_transpose_x_0"), val = bool(false)];
            bool var_5884_transpose_y_0 = const()[name = string("op_5884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_5739_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_5884_cast_fp16 = matmul(transpose_x = var_5884_transpose_x_0, transpose_y = var_5884_transpose_y_0, x = var_5883_cast_fp16, y = v_29_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5884_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_5901_begin_0 = const()[name = string("op_5901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5901_end_0 = const()[name = string("op_5901_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5901_end_mask_0 = const()[name = string("op_5901_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5901_cast_fp16 = slice_by_index(begin = var_5901_begin_0, end = var_5901_end_0, end_mask = var_5901_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5901_cast_fp16")];
            tensor<int32, [4]> var_5916_begin_0 = const()[name = string("op_5916_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5916_end_0 = const()[name = string("op_5916_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5916_end_mask_0 = const()[name = string("op_5916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5916_cast_fp16 = slice_by_index(begin = var_5916_begin_0, end = var_5916_end_0, end_mask = var_5916_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5916_cast_fp16")];
            int32 var_5918 = const()[name = string("op_5918"), val = int32(1)];
            bool mixed_59_interleave_0 = const()[name = string("mixed_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_59_cast_fp16 = concat(axis = var_5918, interleave = mixed_59_interleave_0, values = (var_5901_cast_fp16, var_5916_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_5922_perm_0 = const()[name = string("op_5922_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5926 = const()[name = string("op_5926"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5922_cast_fp16 = transpose(perm = var_5922_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_5926, x = var_5922_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536079936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538177152))))[name = string("model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = add(x = x_169_cast_fp16, y = linear_101_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_5937_promoted_to_fp16 = const()[name = string("op_5937_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5939_cast_fp16 = pow(x = x_179_cast_fp16, y = var_5937_promoted_to_fp16)[name = string("op_5939_cast_fp16")];
            tensor<int32, [1]> var_5941_axes_0 = const()[name = string("op_5941_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5941_keep_dims_0 = const()[name = string("op_5941_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5941_cast_fp16 = reduce_mean(axes = var_5941_axes_0, keep_dims = var_5941_keep_dims_0, x = var_5939_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 var_5942_to_fp16 = const()[name = string("op_5942_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5943_cast_fp16 = add(x = var_5941_cast_fp16, y = var_5942_to_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5944_cast_fp16 = rsqrt(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5945_cast_fp16 = mul(x = x_179_cast_fp16, y = var_5944_cast_fp16)[name = string("op_5945_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538179264)))];
            tensor<fp16, [1, 16, 1024]> input_145_cast_fp16 = mul(x = var_5945_cast_fp16, y = model_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538181376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541327168))))[name = string("model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5950_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_5950_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541333376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544479168))))[name = string("model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_149_cast_fp16 = mul(x = var_5950_cast_fp16, y = linear_103_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544485376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547631168))))[name = string("model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_104_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_181_cast_fp16 = add(x = x_179_cast_fp16, y = linear_104_cast_fp16)[name = string("x_181_cast_fp16")];
            fp16 var_5964_promoted_to_fp16 = const()[name = string("op_5964_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5966_cast_fp16 = pow(x = x_181_cast_fp16, y = var_5964_promoted_to_fp16)[name = string("op_5966_cast_fp16")];
            tensor<int32, [1]> var_5968_axes_0 = const()[name = string("op_5968_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5968_keep_dims_0 = const()[name = string("op_5968_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5968_cast_fp16 = reduce_mean(axes = var_5968_axes_0, keep_dims = var_5968_keep_dims_0, x = var_5966_cast_fp16)[name = string("op_5968_cast_fp16")];
            fp16 var_5969_to_fp16 = const()[name = string("op_5969_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5970_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5969_to_fp16)[name = string("op_5970_cast_fp16")];
            fp32 var_5971_epsilon_0 = const()[name = string("op_5971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5971_cast_fp16 = rsqrt(epsilon = var_5971_epsilon_0, x = var_5970_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5972_cast_fp16 = mul(x = x_181_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5972_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547633280)))];
            tensor<fp16, [1, 16, 1024]> input_151_cast_fp16 = mul(x = var_5972_cast_fp16, y = model_layers_15_input_layernorm_weight_to_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547635392))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549732608))))[name = string("model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_5981 = const()[name = string("op_5981"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5982_cast_fp16 = reshape(shape = var_5981, x = linear_105_cast_fp16)[name = string("op_5982_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5990_promoted_to_fp16 = const()[name = string("op_5990_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5982_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_5992_cast_fp16 = pow(x = x_183_cast_fp16, y = var_5990_promoted_to_fp16)[name = string("op_5992_cast_fp16")];
            tensor<int32, [1]> var_5994_axes_0 = const()[name = string("op_5994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5994_keep_dims_0 = const()[name = string("op_5994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5994_cast_fp16 = reduce_mean(axes = var_5994_axes_0, keep_dims = var_5994_keep_dims_0, x = var_5992_cast_fp16)[name = string("op_5994_cast_fp16")];
            fp16 var_5995_to_fp16 = const()[name = string("op_5995_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5996_cast_fp16 = add(x = var_5994_cast_fp16, y = var_5995_to_fp16)[name = string("op_5996_cast_fp16")];
            fp32 var_5997_epsilon_0 = const()[name = string("op_5997_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5997_cast_fp16 = rsqrt(epsilon = var_5997_epsilon_0, x = var_5996_cast_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5998_cast_fp16 = mul(x = x_183_cast_fp16, y = var_5997_cast_fp16)[name = string("op_5998_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549736768)))];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = mul(x = var_5998_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549737088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550785728))))[name = string("model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6007 = const()[name = string("op_6007"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6008_cast_fp16 = reshape(shape = var_6007, x = linear_106_cast_fp16)[name = string("op_6008_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6016_promoted_to_fp16 = const()[name = string("op_6016_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_6008_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6018_cast_fp16 = pow(x = x_185_cast_fp16, y = var_6016_promoted_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<int32, [1]> var_6020_axes_0 = const()[name = string("op_6020_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6020_keep_dims_0 = const()[name = string("op_6020_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6020_cast_fp16 = reduce_mean(axes = var_6020_axes_0, keep_dims = var_6020_keep_dims_0, x = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp16 var_6021_to_fp16 = const()[name = string("op_6021_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6022_cast_fp16 = add(x = var_6020_cast_fp16, y = var_6021_to_fp16)[name = string("op_6022_cast_fp16")];
            fp32 var_6023_epsilon_0 = const()[name = string("op_6023_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6023_cast_fp16 = rsqrt(epsilon = var_6023_epsilon_0, x = var_6022_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6024_cast_fp16 = mul(x = x_185_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6024_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550787840)))];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = mul(x = var_6024_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550788160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551836800))))[name = string("model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6033 = const()[name = string("op_6033"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6034_cast_fp16 = reshape(shape = var_6033, x = linear_107_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_61_begin_0 = const()[name = string("first_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_61_end_0 = const()[name = string("first_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_61_end_mask_0 = const()[name = string("first_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_61_cast_fp16 = slice_by_index(begin = first_61_begin_0, end = first_61_end_0, end_mask = first_61_end_mask_0, x = x_187_cast_fp16)[name = string("first_61_cast_fp16")];
            tensor<int32, [4]> second_61_begin_0 = const()[name = string("second_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_61_end_0 = const()[name = string("second_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_61_end_mask_0 = const()[name = string("second_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_61_cast_fp16 = slice_by_index(begin = second_61_begin_0, end = second_61_end_0, end_mask = second_61_end_mask_0, x = x_187_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6048_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6049_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6049_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6051_cast_fp16 = sub(x = var_6048_cast_fp16, y = var_6049_cast_fp16)[name = string("op_6051_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6052_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6052_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6053_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6055_cast_fp16 = add(x = var_6052_cast_fp16, y = var_6053_cast_fp16)[name = string("op_6055_cast_fp16")];
            int32 var_6057 = const()[name = string("op_6057"), val = int32(-1)];
            bool q_31_interleave_0 = const()[name = string("q_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_31_cast_fp16 = concat(axis = var_6057, interleave = q_31_interleave_0, values = (var_6051_cast_fp16, var_6055_cast_fp16))[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> first_63_begin_0 = const()[name = string("first_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_63_end_0 = const()[name = string("first_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_63_end_mask_0 = const()[name = string("first_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_63_cast_fp16 = slice_by_index(begin = first_63_begin_0, end = first_63_end_0, end_mask = first_63_end_mask_0, x = x_189_cast_fp16)[name = string("first_63_cast_fp16")];
            tensor<int32, [4]> second_63_begin_0 = const()[name = string("second_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_63_end_0 = const()[name = string("second_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_63_end_mask_0 = const()[name = string("second_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_63_cast_fp16 = slice_by_index(begin = second_63_begin_0, end = second_63_end_0, end_mask = second_63_end_mask_0, x = x_189_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6069_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6070_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6070_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6072_cast_fp16 = sub(x = var_6069_cast_fp16, y = var_6070_cast_fp16)[name = string("op_6072_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6073_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6074_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6076_cast_fp16 = add(x = var_6073_cast_fp16, y = var_6074_cast_fp16)[name = string("op_6076_cast_fp16")];
            int32 var_6078 = const()[name = string("op_6078"), val = int32(-1)];
            bool k_31_interleave_0 = const()[name = string("k_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_31_cast_fp16 = concat(axis = var_6078, interleave = k_31_interleave_0, values = (var_6072_cast_fp16, var_6076_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6089_begin_0 = const()[name = string("op_6089_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6089_end_0 = const()[name = string("op_6089_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6089_end_mask_0 = const()[name = string("op_6089_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6089_stride_0 = const()[name = string("op_6089_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6089_cast_fp16 = slice_by_index(begin = var_6089_begin_0, end = var_6089_end_0, end_mask = var_6089_end_mask_0, stride = var_6089_stride_0, x = q_31_cast_fp16)[name = string("op_6089_cast_fp16")];
            tensor<int32, [4]> var_6099_begin_0 = const()[name = string("op_6099_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6099_end_0 = const()[name = string("op_6099_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6099_end_mask_0 = const()[name = string("op_6099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6099_stride_0 = const()[name = string("op_6099_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6099_cast_fp16 = slice_by_index(begin = var_6099_begin_0, end = var_6099_end_0, end_mask = var_6099_end_mask_0, stride = var_6099_stride_0, x = q_31_cast_fp16)[name = string("op_6099_cast_fp16")];
            int32 var_6101 = const()[name = string("op_6101"), val = int32(2)];
            bool stacked_31_interleave_0 = const()[name = string("stacked_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_31_cast_fp16 = concat(axis = var_6101, interleave = stacked_31_interleave_0, values = (var_6089_cast_fp16, var_6099_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6107_begin_0 = const()[name = string("op_6107_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6107_end_0 = const()[name = string("op_6107_end_0"), val = tensor<int32, [4]>([16, 8, 2064, 128])];
            tensor<bool, [4]> var_6107_end_mask_0 = const()[name = string("op_6107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6107_cast_fp16 = slice_by_index(begin = var_6107_begin_0, end = var_6107_end_0, end_mask = var_6107_end_mask_0, x = read_state_0)[name = string("op_6107_cast_fp16")];
            tensor<int32, [4]> var_6117_begin_0 = const()[name = string("op_6117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6117_end_0 = const()[name = string("op_6117_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6117_end_mask_0 = const()[name = string("op_6117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6117_cast_fp16 = slice_by_index(begin = var_6117_begin_0, end = var_6117_end_0, end_mask = var_6117_end_mask_0, x = var_6107_cast_fp16)[name = string("op_6117_cast_fp16")];
            bool var_6126_transpose_x_1 = const()[name = string("op_6126_transpose_x_1"), val = bool(false)];
            bool var_6126_transpose_y_1 = const()[name = string("op_6126_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6126_cast_fp16 = matmul(transpose_x = var_6126_transpose_x_1, transpose_y = var_6126_transpose_y_1, x = var_6117_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<int32, [4]> var_6129_perm_0 = const()[name = string("op_6129_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6130_to_fp16 = const()[name = string("op_6130_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6129_cast_fp16 = transpose(perm = var_6129_perm_0, x = var_6126_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 2048]> var_6131_cast_fp16 = mul(x = var_6129_cast_fp16, y = var_6130_to_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_31_cast_fp16 = add(x = var_6131_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6137_transpose_x_1 = const()[name = string("op_6137_transpose_x_1"), val = bool(false)];
            bool var_6137_transpose_y_1 = const()[name = string("op_6137_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6137_cast_fp16 = matmul(transpose_x = var_6137_transpose_x_1, transpose_y = var_6137_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6139_cast_fp16 = mul(x = var_6137_cast_fp16, y = var_6138_to_fp16)[name = string("op_6139_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6139_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6143 = const()[name = string("op_6143"), val = int32(-1)];
            bool var_6144_interleave_0 = const()[name = string("op_6144_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6144_cast_fp16 = concat(axis = var_6143, interleave = var_6144_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6144_cast_fp16")];
            int32 var_6145 = const()[name = string("op_6145"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_31_cast_fp16 = softmax(axis = var_6145, x = var_6144_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6152_begin_0 = const()[name = string("op_6152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6152_end_0 = const()[name = string("op_6152_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6152_end_mask_0 = const()[name = string("op_6152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6152_cast_fp16 = slice_by_index(begin = var_6152_begin_0, end = var_6152_end_0, end_mask = var_6152_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<int32, [4]> var_6157_begin_0 = const()[name = string("op_6157_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6157_end_0 = const()[name = string("op_6157_end_0"), val = tensor<int32, [4]>([16, 8, 2064, 128])];
            tensor<bool, [4]> var_6157_end_mask_0 = const()[name = string("op_6157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6157_cast_fp16 = slice_by_index(begin = var_6157_begin_0, end = var_6157_end_0, end_mask = var_6157_end_mask_0, x = read_state_1)[name = string("op_6157_cast_fp16")];
            tensor<int32, [4]> var_6167_begin_0 = const()[name = string("op_6167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6167_end_0 = const()[name = string("op_6167_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6167_end_mask_0 = const()[name = string("op_6167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6167_cast_fp16 = slice_by_index(begin = var_6167_begin_0, end = var_6167_end_0, end_mask = var_6167_end_mask_0, x = var_6157_cast_fp16)[name = string("op_6167_cast_fp16")];
            bool var_6173_transpose_x_0 = const()[name = string("op_6173_transpose_x_0"), val = bool(false)];
            bool var_6173_transpose_y_0 = const()[name = string("op_6173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6173_cast_fp16 = matmul(transpose_x = var_6173_transpose_x_0, transpose_y = var_6173_transpose_y_0, x = var_6152_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6173_cast_fp16")];
            tensor<int32, [4]> var_6178_begin_0 = const()[name = string("op_6178_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6178_end_0 = const()[name = string("op_6178_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6178_end_mask_0 = const()[name = string("op_6178_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6178_cast_fp16 = slice_by_index(begin = var_6178_begin_0, end = var_6178_end_0, end_mask = var_6178_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6178_cast_fp16")];
            bool var_6179_transpose_x_0 = const()[name = string("op_6179_transpose_x_0"), val = bool(false)];
            bool var_6179_transpose_y_0 = const()[name = string("op_6179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6034_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6179_cast_fp16 = matmul(transpose_x = var_6179_transpose_x_0, transpose_y = var_6179_transpose_y_0, x = var_6178_cast_fp16, y = v_31_cast_fp16)[name = string("op_6179_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6173_cast_fp16, y = var_6179_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6196_begin_0 = const()[name = string("op_6196_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6196_end_0 = const()[name = string("op_6196_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6196_end_mask_0 = const()[name = string("op_6196_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6196_cast_fp16 = slice_by_index(begin = var_6196_begin_0, end = var_6196_end_0, end_mask = var_6196_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6196_cast_fp16")];
            tensor<int32, [4]> var_6211_begin_0 = const()[name = string("op_6211_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6211_end_0 = const()[name = string("op_6211_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6211_end_mask_0 = const()[name = string("op_6211_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6211_cast_fp16 = slice_by_index(begin = var_6211_begin_0, end = var_6211_end_0, end_mask = var_6211_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6211_cast_fp16")];
            int32 var_6213 = const()[name = string("op_6213"), val = int32(1)];
            bool mixed_63_interleave_0 = const()[name = string("mixed_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_63_cast_fp16 = concat(axis = var_6213, interleave = mixed_63_interleave_0, values = (var_6196_cast_fp16, var_6211_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6217_perm_0 = const()[name = string("op_6217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6221 = const()[name = string("op_6221"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6217_cast_fp16 = transpose(perm = var_6217_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6221, x = var_6217_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551838912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553936128))))[name = string("model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_108_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = linear_108_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_6232_promoted_to_fp16 = const()[name = string("op_6232_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6234_cast_fp16 = pow(x = x_191_cast_fp16, y = var_6232_promoted_to_fp16)[name = string("op_6234_cast_fp16")];
            tensor<int32, [1]> var_6236_axes_0 = const()[name = string("op_6236_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6236_keep_dims_0 = const()[name = string("op_6236_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6236_cast_fp16 = reduce_mean(axes = var_6236_axes_0, keep_dims = var_6236_keep_dims_0, x = var_6234_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6238_cast_fp16 = add(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            fp32 var_6239_epsilon_0 = const()[name = string("op_6239_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6239_cast_fp16 = rsqrt(epsilon = var_6239_epsilon_0, x = var_6238_cast_fp16)[name = string("op_6239_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6240_cast_fp16 = mul(x = x_191_cast_fp16, y = var_6239_cast_fp16)[name = string("op_6240_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553938240)))];
            tensor<fp16, [1, 16, 1024]> input_155_cast_fp16 = mul(x = var_6240_cast_fp16, y = model_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553940352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557086144))))[name = string("model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6245_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557092352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560238144))))[name = string("model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_159_cast_fp16 = mul(x = var_6245_cast_fp16, y = linear_110_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560244352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563390144))))[name = string("model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_193_cast_fp16 = add(x = x_191_cast_fp16, y = linear_111_cast_fp16)[name = string("x_193_cast_fp16")];
            fp16 var_6259_promoted_to_fp16 = const()[name = string("op_6259_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6261_cast_fp16 = pow(x = x_193_cast_fp16, y = var_6259_promoted_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<int32, [1]> var_6263_axes_0 = const()[name = string("op_6263_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6263_keep_dims_0 = const()[name = string("op_6263_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6263_cast_fp16 = reduce_mean(axes = var_6263_axes_0, keep_dims = var_6263_keep_dims_0, x = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp16 var_6264_to_fp16 = const()[name = string("op_6264_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6265_cast_fp16 = add(x = var_6263_cast_fp16, y = var_6264_to_fp16)[name = string("op_6265_cast_fp16")];
            fp32 var_6266_epsilon_0 = const()[name = string("op_6266_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6266_cast_fp16 = rsqrt(epsilon = var_6266_epsilon_0, x = var_6265_cast_fp16)[name = string("op_6266_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6267_cast_fp16 = mul(x = x_193_cast_fp16, y = var_6266_cast_fp16)[name = string("op_6267_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563392256)))];
            tensor<fp16, [1, 16, 1024]> input_161_cast_fp16 = mul(x = var_6267_cast_fp16, y = model_layers_16_input_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563394368))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565491584))))[name = string("model_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_6276 = const()[name = string("op_6276"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6277_cast_fp16 = reshape(shape = var_6276, x = linear_112_cast_fp16)[name = string("op_6277_cast_fp16")];
            tensor<int32, [4]> x_195_perm_0 = const()[name = string("x_195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6285_promoted_to_fp16 = const()[name = string("op_6285_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_195_cast_fp16 = transpose(perm = x_195_perm_0, x = var_6277_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6287_cast_fp16 = pow(x = x_195_cast_fp16, y = var_6285_promoted_to_fp16)[name = string("op_6287_cast_fp16")];
            tensor<int32, [1]> var_6289_axes_0 = const()[name = string("op_6289_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6289_keep_dims_0 = const()[name = string("op_6289_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6289_cast_fp16 = reduce_mean(axes = var_6289_axes_0, keep_dims = var_6289_keep_dims_0, x = var_6287_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 var_6290_to_fp16 = const()[name = string("op_6290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6291_cast_fp16 = add(x = var_6289_cast_fp16, y = var_6290_to_fp16)[name = string("op_6291_cast_fp16")];
            fp32 var_6292_epsilon_0 = const()[name = string("op_6292_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6292_cast_fp16 = rsqrt(epsilon = var_6292_epsilon_0, x = var_6291_cast_fp16)[name = string("op_6292_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6293_cast_fp16 = mul(x = x_195_cast_fp16, y = var_6292_cast_fp16)[name = string("op_6293_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565495744)))];
            tensor<fp16, [1, 16, 16, 128]> x_199_cast_fp16 = mul(x = var_6293_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565496064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566544704))))[name = string("model_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_6302 = const()[name = string("op_6302"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6303_cast_fp16 = reshape(shape = var_6302, x = linear_113_cast_fp16)[name = string("op_6303_cast_fp16")];
            tensor<int32, [4]> x_197_perm_0 = const()[name = string("x_197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6311_promoted_to_fp16 = const()[name = string("op_6311_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_197_cast_fp16 = transpose(perm = x_197_perm_0, x = var_6303_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6313_cast_fp16 = pow(x = x_197_cast_fp16, y = var_6311_promoted_to_fp16)[name = string("op_6313_cast_fp16")];
            tensor<int32, [1]> var_6315_axes_0 = const()[name = string("op_6315_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6315_keep_dims_0 = const()[name = string("op_6315_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6315_cast_fp16 = reduce_mean(axes = var_6315_axes_0, keep_dims = var_6315_keep_dims_0, x = var_6313_cast_fp16)[name = string("op_6315_cast_fp16")];
            fp16 var_6316_to_fp16 = const()[name = string("op_6316_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6317_cast_fp16 = add(x = var_6315_cast_fp16, y = var_6316_to_fp16)[name = string("op_6317_cast_fp16")];
            fp32 var_6318_epsilon_0 = const()[name = string("op_6318_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6318_cast_fp16 = rsqrt(epsilon = var_6318_epsilon_0, x = var_6317_cast_fp16)[name = string("op_6318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6319_cast_fp16 = mul(x = x_197_cast_fp16, y = var_6318_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566546816)))];
            tensor<fp16, [1, 8, 16, 128]> x_201_cast_fp16 = mul(x = var_6319_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566547136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567595776))))[name = string("model_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_6328 = const()[name = string("op_6328"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6329_cast_fp16 = reshape(shape = var_6328, x = linear_114_cast_fp16)[name = string("op_6329_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_65_begin_0 = const()[name = string("first_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_65_end_0 = const()[name = string("first_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_65_end_mask_0 = const()[name = string("first_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_65_cast_fp16 = slice_by_index(begin = first_65_begin_0, end = first_65_end_0, end_mask = first_65_end_mask_0, x = x_199_cast_fp16)[name = string("first_65_cast_fp16")];
            tensor<int32, [4]> second_65_begin_0 = const()[name = string("second_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_65_end_0 = const()[name = string("second_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_65_end_mask_0 = const()[name = string("second_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_65_cast_fp16 = slice_by_index(begin = second_65_begin_0, end = second_65_end_0, end_mask = second_65_end_mask_0, x = x_199_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6343_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6343_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6344_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6346_cast_fp16 = sub(x = var_6343_cast_fp16, y = var_6344_cast_fp16)[name = string("op_6346_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6347_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6347_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6348_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6350_cast_fp16 = add(x = var_6347_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            int32 var_6352 = const()[name = string("op_6352"), val = int32(-1)];
            bool q_33_interleave_0 = const()[name = string("q_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_33_cast_fp16 = concat(axis = var_6352, interleave = q_33_interleave_0, values = (var_6346_cast_fp16, var_6350_cast_fp16))[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> first_67_begin_0 = const()[name = string("first_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_67_end_0 = const()[name = string("first_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_67_end_mask_0 = const()[name = string("first_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_67_cast_fp16 = slice_by_index(begin = first_67_begin_0, end = first_67_end_0, end_mask = first_67_end_mask_0, x = x_201_cast_fp16)[name = string("first_67_cast_fp16")];
            tensor<int32, [4]> second_67_begin_0 = const()[name = string("second_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_67_end_0 = const()[name = string("second_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_67_end_mask_0 = const()[name = string("second_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_67_cast_fp16 = slice_by_index(begin = second_67_begin_0, end = second_67_end_0, end_mask = second_67_end_mask_0, x = x_201_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6364_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6365_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6365_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6367_cast_fp16 = sub(x = var_6364_cast_fp16, y = var_6365_cast_fp16)[name = string("op_6367_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6368_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6369_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6371_cast_fp16 = add(x = var_6368_cast_fp16, y = var_6369_cast_fp16)[name = string("op_6371_cast_fp16")];
            int32 var_6373 = const()[name = string("op_6373"), val = int32(-1)];
            bool k_33_interleave_0 = const()[name = string("k_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_33_cast_fp16 = concat(axis = var_6373, interleave = k_33_interleave_0, values = (var_6367_cast_fp16, var_6371_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6384_begin_0 = const()[name = string("op_6384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6384_end_0 = const()[name = string("op_6384_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6384_end_mask_0 = const()[name = string("op_6384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6384_stride_0 = const()[name = string("op_6384_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6384_cast_fp16 = slice_by_index(begin = var_6384_begin_0, end = var_6384_end_0, end_mask = var_6384_end_mask_0, stride = var_6384_stride_0, x = q_33_cast_fp16)[name = string("op_6384_cast_fp16")];
            tensor<int32, [4]> var_6394_begin_0 = const()[name = string("op_6394_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6394_end_0 = const()[name = string("op_6394_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6394_end_mask_0 = const()[name = string("op_6394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6394_stride_0 = const()[name = string("op_6394_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6394_cast_fp16 = slice_by_index(begin = var_6394_begin_0, end = var_6394_end_0, end_mask = var_6394_end_mask_0, stride = var_6394_stride_0, x = q_33_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6396 = const()[name = string("op_6396"), val = int32(2)];
            bool stacked_33_interleave_0 = const()[name = string("stacked_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_33_cast_fp16 = concat(axis = var_6396, interleave = stacked_33_interleave_0, values = (var_6384_cast_fp16, var_6394_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6402_begin_0 = const()[name = string("op_6402_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6402_end_0 = const()[name = string("op_6402_end_0"), val = tensor<int32, [4]>([17, 8, 2064, 128])];
            tensor<bool, [4]> var_6402_end_mask_0 = const()[name = string("op_6402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6402_cast_fp16 = slice_by_index(begin = var_6402_begin_0, end = var_6402_end_0, end_mask = var_6402_end_mask_0, x = read_state_0)[name = string("op_6402_cast_fp16")];
            tensor<int32, [4]> var_6412_begin_0 = const()[name = string("op_6412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6412_end_0 = const()[name = string("op_6412_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6412_end_mask_0 = const()[name = string("op_6412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6412_cast_fp16 = slice_by_index(begin = var_6412_begin_0, end = var_6412_end_0, end_mask = var_6412_end_mask_0, x = var_6402_cast_fp16)[name = string("op_6412_cast_fp16")];
            bool var_6421_transpose_x_1 = const()[name = string("op_6421_transpose_x_1"), val = bool(false)];
            bool var_6421_transpose_y_1 = const()[name = string("op_6421_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6421_cast_fp16 = matmul(transpose_x = var_6421_transpose_x_1, transpose_y = var_6421_transpose_y_1, x = var_6412_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<int32, [4]> var_6424_perm_0 = const()[name = string("op_6424_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6425_to_fp16 = const()[name = string("op_6425_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6424_cast_fp16 = transpose(perm = var_6424_perm_0, x = var_6421_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 2048]> var_6426_cast_fp16 = mul(x = var_6424_cast_fp16, y = var_6425_to_fp16)[name = string("op_6426_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_33_cast_fp16 = add(x = var_6426_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6432_transpose_x_1 = const()[name = string("op_6432_transpose_x_1"), val = bool(false)];
            bool var_6432_transpose_y_1 = const()[name = string("op_6432_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6432_cast_fp16 = matmul(transpose_x = var_6432_transpose_x_1, transpose_y = var_6432_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6432_cast_fp16")];
            fp16 var_6433_to_fp16 = const()[name = string("op_6433_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6434_cast_fp16 = mul(x = var_6432_cast_fp16, y = var_6433_to_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6434_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6438 = const()[name = string("op_6438"), val = int32(-1)];
            bool var_6439_interleave_0 = const()[name = string("op_6439_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6439_cast_fp16 = concat(axis = var_6438, interleave = var_6439_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6439_cast_fp16")];
            int32 var_6440 = const()[name = string("op_6440"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_33_cast_fp16 = softmax(axis = var_6440, x = var_6439_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6447_begin_0 = const()[name = string("op_6447_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6447_end_0 = const()[name = string("op_6447_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6447_end_mask_0 = const()[name = string("op_6447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6447_cast_fp16 = slice_by_index(begin = var_6447_begin_0, end = var_6447_end_0, end_mask = var_6447_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6447_cast_fp16")];
            tensor<int32, [4]> var_6452_begin_0 = const()[name = string("op_6452_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6452_end_0 = const()[name = string("op_6452_end_0"), val = tensor<int32, [4]>([17, 8, 2064, 128])];
            tensor<bool, [4]> var_6452_end_mask_0 = const()[name = string("op_6452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6452_cast_fp16 = slice_by_index(begin = var_6452_begin_0, end = var_6452_end_0, end_mask = var_6452_end_mask_0, x = read_state_1)[name = string("op_6452_cast_fp16")];
            tensor<int32, [4]> var_6462_begin_0 = const()[name = string("op_6462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6462_end_0 = const()[name = string("op_6462_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6462_end_mask_0 = const()[name = string("op_6462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6462_cast_fp16 = slice_by_index(begin = var_6462_begin_0, end = var_6462_end_0, end_mask = var_6462_end_mask_0, x = var_6452_cast_fp16)[name = string("op_6462_cast_fp16")];
            bool var_6468_transpose_x_0 = const()[name = string("op_6468_transpose_x_0"), val = bool(false)];
            bool var_6468_transpose_y_0 = const()[name = string("op_6468_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6468_cast_fp16 = matmul(transpose_x = var_6468_transpose_x_0, transpose_y = var_6468_transpose_y_0, x = var_6447_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6468_cast_fp16")];
            tensor<int32, [4]> var_6473_begin_0 = const()[name = string("op_6473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6473_end_0 = const()[name = string("op_6473_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6473_end_mask_0 = const()[name = string("op_6473_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6473_cast_fp16 = slice_by_index(begin = var_6473_begin_0, end = var_6473_end_0, end_mask = var_6473_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6473_cast_fp16")];
            bool var_6474_transpose_x_0 = const()[name = string("op_6474_transpose_x_0"), val = bool(false)];
            bool var_6474_transpose_y_0 = const()[name = string("op_6474_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6329_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6474_cast_fp16 = matmul(transpose_x = var_6474_transpose_x_0, transpose_y = var_6474_transpose_y_0, x = var_6473_cast_fp16, y = v_33_cast_fp16)[name = string("op_6474_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6468_cast_fp16, y = var_6474_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6491_begin_0 = const()[name = string("op_6491_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6491_end_0 = const()[name = string("op_6491_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6491_end_mask_0 = const()[name = string("op_6491_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6491_cast_fp16 = slice_by_index(begin = var_6491_begin_0, end = var_6491_end_0, end_mask = var_6491_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6491_cast_fp16")];
            tensor<int32, [4]> var_6506_begin_0 = const()[name = string("op_6506_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6506_end_0 = const()[name = string("op_6506_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6506_end_mask_0 = const()[name = string("op_6506_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6506_cast_fp16 = slice_by_index(begin = var_6506_begin_0, end = var_6506_end_0, end_mask = var_6506_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6506_cast_fp16")];
            int32 var_6508 = const()[name = string("op_6508"), val = int32(1)];
            bool mixed_67_interleave_0 = const()[name = string("mixed_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_67_cast_fp16 = concat(axis = var_6508, interleave = mixed_67_interleave_0, values = (var_6491_cast_fp16, var_6506_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6512_perm_0 = const()[name = string("op_6512_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6516 = const()[name = string("op_6516"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6512_cast_fp16 = transpose(perm = var_6512_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6516, x = var_6512_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567597888))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569695104))))[name = string("model_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = add(x = x_193_cast_fp16, y = linear_115_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_6527_promoted_to_fp16 = const()[name = string("op_6527_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6529_cast_fp16 = pow(x = x_203_cast_fp16, y = var_6527_promoted_to_fp16)[name = string("op_6529_cast_fp16")];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6531_keep_dims_0 = const()[name = string("op_6531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6531_cast_fp16 = reduce_mean(axes = var_6531_axes_0, keep_dims = var_6531_keep_dims_0, x = var_6529_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6533_cast_fp16 = add(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("op_6533_cast_fp16")];
            fp32 var_6534_epsilon_0 = const()[name = string("op_6534_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6534_cast_fp16 = rsqrt(epsilon = var_6534_epsilon_0, x = var_6533_cast_fp16)[name = string("op_6534_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6535_cast_fp16 = mul(x = x_203_cast_fp16, y = var_6534_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569697216)))];
            tensor<fp16, [1, 16, 1024]> input_165_cast_fp16 = mul(x = var_6535_cast_fp16, y = model_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569699328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572845120))))[name = string("model_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6540_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572851328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575997120))))[name = string("model_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_169_cast_fp16 = mul(x = var_6540_cast_fp16, y = linear_117_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576003328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579149120))))[name = string("model_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_205_cast_fp16 = add(x = x_203_cast_fp16, y = linear_118_cast_fp16)[name = string("x_205_cast_fp16")];
            fp16 var_6554_promoted_to_fp16 = const()[name = string("op_6554_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6556_cast_fp16 = pow(x = x_205_cast_fp16, y = var_6554_promoted_to_fp16)[name = string("op_6556_cast_fp16")];
            tensor<int32, [1]> var_6558_axes_0 = const()[name = string("op_6558_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6558_keep_dims_0 = const()[name = string("op_6558_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6558_cast_fp16 = reduce_mean(axes = var_6558_axes_0, keep_dims = var_6558_keep_dims_0, x = var_6556_cast_fp16)[name = string("op_6558_cast_fp16")];
            fp16 var_6559_to_fp16 = const()[name = string("op_6559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6560_cast_fp16 = add(x = var_6558_cast_fp16, y = var_6559_to_fp16)[name = string("op_6560_cast_fp16")];
            fp32 var_6561_epsilon_0 = const()[name = string("op_6561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6561_cast_fp16 = rsqrt(epsilon = var_6561_epsilon_0, x = var_6560_cast_fp16)[name = string("op_6561_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6562_cast_fp16 = mul(x = x_205_cast_fp16, y = var_6561_cast_fp16)[name = string("op_6562_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579151232)))];
            tensor<fp16, [1, 16, 1024]> input_171_cast_fp16 = mul(x = var_6562_cast_fp16, y = model_layers_17_input_layernorm_weight_to_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579153344))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581250560))))[name = string("model_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6571 = const()[name = string("op_6571"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6572_cast_fp16 = reshape(shape = var_6571, x = linear_119_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<int32, [4]> x_207_perm_0 = const()[name = string("x_207_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6580_promoted_to_fp16 = const()[name = string("op_6580_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = transpose(perm = x_207_perm_0, x = var_6572_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_6582_cast_fp16 = pow(x = x_207_cast_fp16, y = var_6580_promoted_to_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6584_axes_0 = const()[name = string("op_6584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6584_keep_dims_0 = const()[name = string("op_6584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6584_cast_fp16 = reduce_mean(axes = var_6584_axes_0, keep_dims = var_6584_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6584_cast_fp16")];
            fp16 var_6585_to_fp16 = const()[name = string("op_6585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6586_cast_fp16 = add(x = var_6584_cast_fp16, y = var_6585_to_fp16)[name = string("op_6586_cast_fp16")];
            fp32 var_6587_epsilon_0 = const()[name = string("op_6587_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6587_cast_fp16 = rsqrt(epsilon = var_6587_epsilon_0, x = var_6586_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6588_cast_fp16 = mul(x = x_207_cast_fp16, y = var_6587_cast_fp16)[name = string("op_6588_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581254720)))];
            tensor<fp16, [1, 16, 16, 128]> x_211_cast_fp16 = mul(x = var_6588_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_211_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581255040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582303680))))[name = string("model_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6597 = const()[name = string("op_6597"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6598_cast_fp16 = reshape(shape = var_6597, x = linear_120_cast_fp16)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6606_promoted_to_fp16 = const()[name = string("op_6606_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_6598_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_6608_cast_fp16 = pow(x = x_209_cast_fp16, y = var_6606_promoted_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<int32, [1]> var_6610_axes_0 = const()[name = string("op_6610_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6610_keep_dims_0 = const()[name = string("op_6610_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6610_cast_fp16 = reduce_mean(axes = var_6610_axes_0, keep_dims = var_6610_keep_dims_0, x = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp16 var_6611_to_fp16 = const()[name = string("op_6611_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6612_cast_fp16 = add(x = var_6610_cast_fp16, y = var_6611_to_fp16)[name = string("op_6612_cast_fp16")];
            fp32 var_6613_epsilon_0 = const()[name = string("op_6613_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6613_cast_fp16 = rsqrt(epsilon = var_6613_epsilon_0, x = var_6612_cast_fp16)[name = string("op_6613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6614_cast_fp16 = mul(x = x_209_cast_fp16, y = var_6613_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582305792)))];
            tensor<fp16, [1, 8, 16, 128]> x_213_cast_fp16 = mul(x = var_6614_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582306112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583354752))))[name = string("model_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_6623 = const()[name = string("op_6623"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6624_cast_fp16 = reshape(shape = var_6623, x = linear_121_cast_fp16)[name = string("op_6624_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_69_begin_0 = const()[name = string("first_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_69_end_0 = const()[name = string("first_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_69_end_mask_0 = const()[name = string("first_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_69_cast_fp16 = slice_by_index(begin = first_69_begin_0, end = first_69_end_0, end_mask = first_69_end_mask_0, x = x_211_cast_fp16)[name = string("first_69_cast_fp16")];
            tensor<int32, [4]> second_69_begin_0 = const()[name = string("second_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_69_end_0 = const()[name = string("second_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_69_end_mask_0 = const()[name = string("second_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_69_cast_fp16 = slice_by_index(begin = second_69_begin_0, end = second_69_end_0, end_mask = second_69_end_mask_0, x = x_211_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6638_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6638_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6639_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6639_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6641_cast_fp16 = sub(x = var_6638_cast_fp16, y = var_6639_cast_fp16)[name = string("op_6641_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6642_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6642_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6643_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6645_cast_fp16 = add(x = var_6642_cast_fp16, y = var_6643_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6647 = const()[name = string("op_6647"), val = int32(-1)];
            bool q_35_interleave_0 = const()[name = string("q_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_35_cast_fp16 = concat(axis = var_6647, interleave = q_35_interleave_0, values = (var_6641_cast_fp16, var_6645_cast_fp16))[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> first_71_begin_0 = const()[name = string("first_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_71_end_0 = const()[name = string("first_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_71_end_mask_0 = const()[name = string("first_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_71_cast_fp16 = slice_by_index(begin = first_71_begin_0, end = first_71_end_0, end_mask = first_71_end_mask_0, x = x_213_cast_fp16)[name = string("first_71_cast_fp16")];
            tensor<int32, [4]> second_71_begin_0 = const()[name = string("second_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_71_end_0 = const()[name = string("second_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_71_end_mask_0 = const()[name = string("second_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_71_cast_fp16 = slice_by_index(begin = second_71_begin_0, end = second_71_end_0, end_mask = second_71_end_mask_0, x = x_213_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6659_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6659_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6660_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6660_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6662_cast_fp16 = sub(x = var_6659_cast_fp16, y = var_6660_cast_fp16)[name = string("op_6662_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6663_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6664_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6664_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6666_cast_fp16 = add(x = var_6663_cast_fp16, y = var_6664_cast_fp16)[name = string("op_6666_cast_fp16")];
            int32 var_6668 = const()[name = string("op_6668"), val = int32(-1)];
            bool k_35_interleave_0 = const()[name = string("k_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_35_cast_fp16 = concat(axis = var_6668, interleave = k_35_interleave_0, values = (var_6662_cast_fp16, var_6666_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_6679_begin_0 = const()[name = string("op_6679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6679_end_0 = const()[name = string("op_6679_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6679_end_mask_0 = const()[name = string("op_6679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6679_stride_0 = const()[name = string("op_6679_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6679_cast_fp16 = slice_by_index(begin = var_6679_begin_0, end = var_6679_end_0, end_mask = var_6679_end_mask_0, stride = var_6679_stride_0, x = q_35_cast_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [4]> var_6689_begin_0 = const()[name = string("op_6689_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6689_end_0 = const()[name = string("op_6689_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6689_end_mask_0 = const()[name = string("op_6689_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6689_stride_0 = const()[name = string("op_6689_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6689_cast_fp16 = slice_by_index(begin = var_6689_begin_0, end = var_6689_end_0, end_mask = var_6689_end_mask_0, stride = var_6689_stride_0, x = q_35_cast_fp16)[name = string("op_6689_cast_fp16")];
            int32 var_6691 = const()[name = string("op_6691"), val = int32(2)];
            bool stacked_35_interleave_0 = const()[name = string("stacked_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_35_cast_fp16 = concat(axis = var_6691, interleave = stacked_35_interleave_0, values = (var_6679_cast_fp16, var_6689_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([18, 8, 2064, 128])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = read_state_0)[name = string("op_6697_cast_fp16")];
            tensor<int32, [4]> var_6707_begin_0 = const()[name = string("op_6707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6707_end_0 = const()[name = string("op_6707_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6707_end_mask_0 = const()[name = string("op_6707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6707_cast_fp16 = slice_by_index(begin = var_6707_begin_0, end = var_6707_end_0, end_mask = var_6707_end_mask_0, x = var_6697_cast_fp16)[name = string("op_6707_cast_fp16")];
            bool var_6716_transpose_x_1 = const()[name = string("op_6716_transpose_x_1"), val = bool(false)];
            bool var_6716_transpose_y_1 = const()[name = string("op_6716_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_6716_cast_fp16 = matmul(transpose_x = var_6716_transpose_x_1, transpose_y = var_6716_transpose_y_1, x = var_6707_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_6716_cast_fp16")];
            tensor<int32, [4]> var_6719_perm_0 = const()[name = string("op_6719_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6720_to_fp16 = const()[name = string("op_6720_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_6719_cast_fp16 = transpose(perm = var_6719_perm_0, x = var_6716_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 2048]> var_6721_cast_fp16 = mul(x = var_6719_cast_fp16, y = var_6720_to_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_35_cast_fp16 = add(x = var_6721_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_6727_transpose_x_1 = const()[name = string("op_6727_transpose_x_1"), val = bool(false)];
            bool var_6727_transpose_y_1 = const()[name = string("op_6727_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6727_cast_fp16 = matmul(transpose_x = var_6727_transpose_x_1, transpose_y = var_6727_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_6727_cast_fp16")];
            fp16 var_6728_to_fp16 = const()[name = string("op_6728_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6729_cast_fp16 = mul(x = var_6727_cast_fp16, y = var_6728_to_fp16)[name = string("op_6729_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_6729_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_6733 = const()[name = string("op_6733"), val = int32(-1)];
            bool var_6734_interleave_0 = const()[name = string("op_6734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_6734_cast_fp16 = concat(axis = var_6733, interleave = var_6734_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_6734_cast_fp16")];
            int32 var_6735 = const()[name = string("op_6735"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_35_cast_fp16 = softmax(axis = var_6735, x = var_6734_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_6742_begin_0 = const()[name = string("op_6742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6742_end_0 = const()[name = string("op_6742_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_6742_end_mask_0 = const()[name = string("op_6742_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_6742_cast_fp16 = slice_by_index(begin = var_6742_begin_0, end = var_6742_end_0, end_mask = var_6742_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [4]> var_6747_begin_0 = const()[name = string("op_6747_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6747_end_0 = const()[name = string("op_6747_end_0"), val = tensor<int32, [4]>([18, 8, 2064, 128])];
            tensor<bool, [4]> var_6747_end_mask_0 = const()[name = string("op_6747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6747_cast_fp16 = slice_by_index(begin = var_6747_begin_0, end = var_6747_end_0, end_mask = var_6747_end_mask_0, x = read_state_1)[name = string("op_6747_cast_fp16")];
            tensor<int32, [4]> var_6757_begin_0 = const()[name = string("op_6757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6757_end_0 = const()[name = string("op_6757_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_6757_end_mask_0 = const()[name = string("op_6757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_6757_cast_fp16 = slice_by_index(begin = var_6757_begin_0, end = var_6757_end_0, end_mask = var_6757_end_mask_0, x = var_6747_cast_fp16)[name = string("op_6757_cast_fp16")];
            bool var_6763_transpose_x_0 = const()[name = string("op_6763_transpose_x_0"), val = bool(false)];
            bool var_6763_transpose_y_0 = const()[name = string("op_6763_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6763_cast_fp16 = matmul(transpose_x = var_6763_transpose_x_0, transpose_y = var_6763_transpose_y_0, x = var_6742_cast_fp16, y = var_6757_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<int32, [4]> var_6768_begin_0 = const()[name = string("op_6768_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_6768_end_0 = const()[name = string("op_6768_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_6768_end_mask_0 = const()[name = string("op_6768_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6768_cast_fp16 = slice_by_index(begin = var_6768_begin_0, end = var_6768_end_0, end_mask = var_6768_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6768_cast_fp16")];
            bool var_6769_transpose_x_0 = const()[name = string("op_6769_transpose_x_0"), val = bool(false)];
            bool var_6769_transpose_y_0 = const()[name = string("op_6769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_6624_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_6769_cast_fp16 = matmul(transpose_x = var_6769_transpose_x_0, transpose_y = var_6769_transpose_y_0, x = var_6768_cast_fp16, y = v_35_cast_fp16)[name = string("op_6769_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_6763_cast_fp16, y = var_6769_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_6786_begin_0 = const()[name = string("op_6786_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6786_end_0 = const()[name = string("op_6786_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6786_end_mask_0 = const()[name = string("op_6786_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6786_cast_fp16 = slice_by_index(begin = var_6786_begin_0, end = var_6786_end_0, end_mask = var_6786_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<int32, [4]> var_6801_begin_0 = const()[name = string("op_6801_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6801_end_0 = const()[name = string("op_6801_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6801_end_mask_0 = const()[name = string("op_6801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6801_cast_fp16 = slice_by_index(begin = var_6801_begin_0, end = var_6801_end_0, end_mask = var_6801_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6801_cast_fp16")];
            int32 var_6803 = const()[name = string("op_6803"), val = int32(1)];
            bool mixed_71_interleave_0 = const()[name = string("mixed_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_71_cast_fp16 = concat(axis = var_6803, interleave = mixed_71_interleave_0, values = (var_6786_cast_fp16, var_6801_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_6807_perm_0 = const()[name = string("op_6807_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6811 = const()[name = string("op_6811"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6807_cast_fp16 = transpose(perm = var_6807_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_6811, x = var_6807_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583356864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585454080))))[name = string("model_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_122_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_215_cast_fp16 = add(x = x_205_cast_fp16, y = linear_122_cast_fp16)[name = string("x_215_cast_fp16")];
            fp16 var_6822_promoted_to_fp16 = const()[name = string("op_6822_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6824_cast_fp16 = pow(x = x_215_cast_fp16, y = var_6822_promoted_to_fp16)[name = string("op_6824_cast_fp16")];
            tensor<int32, [1]> var_6826_axes_0 = const()[name = string("op_6826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6826_keep_dims_0 = const()[name = string("op_6826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6826_cast_fp16 = reduce_mean(axes = var_6826_axes_0, keep_dims = var_6826_keep_dims_0, x = var_6824_cast_fp16)[name = string("op_6826_cast_fp16")];
            fp16 var_6827_to_fp16 = const()[name = string("op_6827_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6828_cast_fp16 = add(x = var_6826_cast_fp16, y = var_6827_to_fp16)[name = string("op_6828_cast_fp16")];
            fp32 var_6829_epsilon_0 = const()[name = string("op_6829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6829_cast_fp16 = rsqrt(epsilon = var_6829_epsilon_0, x = var_6828_cast_fp16)[name = string("op_6829_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6830_cast_fp16 = mul(x = x_215_cast_fp16, y = var_6829_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585456192)))];
            tensor<fp16, [1, 16, 1024]> input_175_cast_fp16 = mul(x = var_6830_cast_fp16, y = model_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585458304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588604096))))[name = string("model_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6835_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_6835_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588610304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591756096))))[name = string("model_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_179_cast_fp16 = mul(x = var_6835_cast_fp16, y = linear_124_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591762304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594908096))))[name = string("model_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_217_cast_fp16 = add(x = x_215_cast_fp16, y = linear_125_cast_fp16)[name = string("x_217_cast_fp16")];
            fp16 var_6849_promoted_to_fp16 = const()[name = string("op_6849_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6851_cast_fp16 = pow(x = x_217_cast_fp16, y = var_6849_promoted_to_fp16)[name = string("op_6851_cast_fp16")];
            tensor<int32, [1]> var_6853_axes_0 = const()[name = string("op_6853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6853_keep_dims_0 = const()[name = string("op_6853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6853_cast_fp16 = reduce_mean(axes = var_6853_axes_0, keep_dims = var_6853_keep_dims_0, x = var_6851_cast_fp16)[name = string("op_6853_cast_fp16")];
            fp16 var_6854_to_fp16 = const()[name = string("op_6854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6855_cast_fp16 = add(x = var_6853_cast_fp16, y = var_6854_to_fp16)[name = string("op_6855_cast_fp16")];
            fp32 var_6856_epsilon_0 = const()[name = string("op_6856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6856_cast_fp16 = rsqrt(epsilon = var_6856_epsilon_0, x = var_6855_cast_fp16)[name = string("op_6856_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6857_cast_fp16 = mul(x = x_217_cast_fp16, y = var_6856_cast_fp16)[name = string("op_6857_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594910208)))];
            tensor<fp16, [1, 16, 1024]> input_181_cast_fp16 = mul(x = var_6857_cast_fp16, y = model_layers_18_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594912320))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597009536))))[name = string("model_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_6866 = const()[name = string("op_6866"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6867_cast_fp16 = reshape(shape = var_6866, x = linear_126_cast_fp16)[name = string("op_6867_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6875_promoted_to_fp16 = const()[name = string("op_6875_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6867_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_6877_cast_fp16 = pow(x = x_219_cast_fp16, y = var_6875_promoted_to_fp16)[name = string("op_6877_cast_fp16")];
            tensor<int32, [1]> var_6879_axes_0 = const()[name = string("op_6879_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6879_keep_dims_0 = const()[name = string("op_6879_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6879_cast_fp16 = reduce_mean(axes = var_6879_axes_0, keep_dims = var_6879_keep_dims_0, x = var_6877_cast_fp16)[name = string("op_6879_cast_fp16")];
            fp16 var_6880_to_fp16 = const()[name = string("op_6880_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6881_cast_fp16 = add(x = var_6879_cast_fp16, y = var_6880_to_fp16)[name = string("op_6881_cast_fp16")];
            fp32 var_6882_epsilon_0 = const()[name = string("op_6882_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6882_cast_fp16 = rsqrt(epsilon = var_6882_epsilon_0, x = var_6881_cast_fp16)[name = string("op_6882_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6883_cast_fp16 = mul(x = x_219_cast_fp16, y = var_6882_cast_fp16)[name = string("op_6883_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597013696)))];
            tensor<fp16, [1, 16, 16, 128]> x_223_cast_fp16 = mul(x = var_6883_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597014016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598062656))))[name = string("model_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6892 = const()[name = string("op_6892"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6893_cast_fp16 = reshape(shape = var_6892, x = linear_127_cast_fp16)[name = string("op_6893_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6901_promoted_to_fp16 = const()[name = string("op_6901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6893_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_6903_cast_fp16 = pow(x = x_221_cast_fp16, y = var_6901_promoted_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<int32, [1]> var_6905_axes_0 = const()[name = string("op_6905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6905_keep_dims_0 = const()[name = string("op_6905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6905_cast_fp16 = reduce_mean(axes = var_6905_axes_0, keep_dims = var_6905_keep_dims_0, x = var_6903_cast_fp16)[name = string("op_6905_cast_fp16")];
            fp16 var_6906_to_fp16 = const()[name = string("op_6906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6907_cast_fp16 = add(x = var_6905_cast_fp16, y = var_6906_to_fp16)[name = string("op_6907_cast_fp16")];
            fp32 var_6908_epsilon_0 = const()[name = string("op_6908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6908_cast_fp16 = rsqrt(epsilon = var_6908_epsilon_0, x = var_6907_cast_fp16)[name = string("op_6908_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6909_cast_fp16 = mul(x = x_221_cast_fp16, y = var_6908_cast_fp16)[name = string("op_6909_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598064768)))];
            tensor<fp16, [1, 8, 16, 128]> x_225_cast_fp16 = mul(x = var_6909_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598065088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599113728))))[name = string("model_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6918 = const()[name = string("op_6918"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6919_cast_fp16 = reshape(shape = var_6918, x = linear_128_cast_fp16)[name = string("op_6919_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_73_begin_0 = const()[name = string("first_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_73_end_0 = const()[name = string("first_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_73_end_mask_0 = const()[name = string("first_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_73_cast_fp16 = slice_by_index(begin = first_73_begin_0, end = first_73_end_0, end_mask = first_73_end_mask_0, x = x_223_cast_fp16)[name = string("first_73_cast_fp16")];
            tensor<int32, [4]> second_73_begin_0 = const()[name = string("second_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_73_end_0 = const()[name = string("second_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_73_end_mask_0 = const()[name = string("second_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_73_cast_fp16 = slice_by_index(begin = second_73_begin_0, end = second_73_end_0, end_mask = second_73_end_mask_0, x = x_223_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6933_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6933_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6934_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6934_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6936_cast_fp16 = sub(x = var_6933_cast_fp16, y = var_6934_cast_fp16)[name = string("op_6936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6937_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6938_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6938_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6940_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6938_cast_fp16)[name = string("op_6940_cast_fp16")];
            int32 var_6942 = const()[name = string("op_6942"), val = int32(-1)];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_37_cast_fp16 = concat(axis = var_6942, interleave = q_37_interleave_0, values = (var_6936_cast_fp16, var_6940_cast_fp16))[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> first_75_begin_0 = const()[name = string("first_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_75_end_0 = const()[name = string("first_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_75_end_mask_0 = const()[name = string("first_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_75_cast_fp16 = slice_by_index(begin = first_75_begin_0, end = first_75_end_0, end_mask = first_75_end_mask_0, x = x_225_cast_fp16)[name = string("first_75_cast_fp16")];
            tensor<int32, [4]> second_75_begin_0 = const()[name = string("second_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_75_end_0 = const()[name = string("second_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_75_end_mask_0 = const()[name = string("second_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_75_cast_fp16 = slice_by_index(begin = second_75_begin_0, end = second_75_end_0, end_mask = second_75_end_mask_0, x = x_225_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6954_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6954_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6955_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6955_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6957_cast_fp16 = sub(x = var_6954_cast_fp16, y = var_6955_cast_fp16)[name = string("op_6957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6958_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6959_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6961_cast_fp16 = add(x = var_6958_cast_fp16, y = var_6959_cast_fp16)[name = string("op_6961_cast_fp16")];
            int32 var_6963 = const()[name = string("op_6963"), val = int32(-1)];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_37_cast_fp16 = concat(axis = var_6963, interleave = k_37_interleave_0, values = (var_6957_cast_fp16, var_6961_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_6974_begin_0 = const()[name = string("op_6974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6974_end_0 = const()[name = string("op_6974_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6974_end_mask_0 = const()[name = string("op_6974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6974_stride_0 = const()[name = string("op_6974_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6974_cast_fp16 = slice_by_index(begin = var_6974_begin_0, end = var_6974_end_0, end_mask = var_6974_end_mask_0, stride = var_6974_stride_0, x = q_37_cast_fp16)[name = string("op_6974_cast_fp16")];
            tensor<int32, [4]> var_6984_begin_0 = const()[name = string("op_6984_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6984_end_0 = const()[name = string("op_6984_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6984_end_mask_0 = const()[name = string("op_6984_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6984_stride_0 = const()[name = string("op_6984_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6984_cast_fp16 = slice_by_index(begin = var_6984_begin_0, end = var_6984_end_0, end_mask = var_6984_end_mask_0, stride = var_6984_stride_0, x = q_37_cast_fp16)[name = string("op_6984_cast_fp16")];
            int32 var_6986 = const()[name = string("op_6986"), val = int32(2)];
            bool stacked_37_interleave_0 = const()[name = string("stacked_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_37_cast_fp16 = concat(axis = var_6986, interleave = stacked_37_interleave_0, values = (var_6974_cast_fp16, var_6984_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_6992_begin_0 = const()[name = string("op_6992_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_6992_end_0 = const()[name = string("op_6992_end_0"), val = tensor<int32, [4]>([19, 8, 2064, 128])];
            tensor<bool, [4]> var_6992_end_mask_0 = const()[name = string("op_6992_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_6992_cast_fp16 = slice_by_index(begin = var_6992_begin_0, end = var_6992_end_0, end_mask = var_6992_end_mask_0, x = read_state_0)[name = string("op_6992_cast_fp16")];
            tensor<int32, [4]> var_7002_begin_0 = const()[name = string("op_7002_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7002_end_0 = const()[name = string("op_7002_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7002_end_mask_0 = const()[name = string("op_7002_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7002_cast_fp16 = slice_by_index(begin = var_7002_begin_0, end = var_7002_end_0, end_mask = var_7002_end_mask_0, x = var_6992_cast_fp16)[name = string("op_7002_cast_fp16")];
            bool var_7011_transpose_x_1 = const()[name = string("op_7011_transpose_x_1"), val = bool(false)];
            bool var_7011_transpose_y_1 = const()[name = string("op_7011_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7011_cast_fp16 = matmul(transpose_x = var_7011_transpose_x_1, transpose_y = var_7011_transpose_y_1, x = var_7002_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<int32, [4]> var_7014_perm_0 = const()[name = string("op_7014_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7015_to_fp16 = const()[name = string("op_7015_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7014_cast_fp16 = transpose(perm = var_7014_perm_0, x = var_7011_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 2048]> var_7016_cast_fp16 = mul(x = var_7014_cast_fp16, y = var_7015_to_fp16)[name = string("op_7016_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_37_cast_fp16 = add(x = var_7016_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7022_transpose_x_1 = const()[name = string("op_7022_transpose_x_1"), val = bool(false)];
            bool var_7022_transpose_y_1 = const()[name = string("op_7022_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7022_cast_fp16 = matmul(transpose_x = var_7022_transpose_x_1, transpose_y = var_7022_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7022_cast_fp16")];
            fp16 var_7023_to_fp16 = const()[name = string("op_7023_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7024_cast_fp16 = mul(x = var_7022_cast_fp16, y = var_7023_to_fp16)[name = string("op_7024_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7024_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7028 = const()[name = string("op_7028"), val = int32(-1)];
            bool var_7029_interleave_0 = const()[name = string("op_7029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7029_cast_fp16 = concat(axis = var_7028, interleave = var_7029_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7029_cast_fp16")];
            int32 var_7030 = const()[name = string("op_7030"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_37_cast_fp16 = softmax(axis = var_7030, x = var_7029_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7037_begin_0 = const()[name = string("op_7037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7037_end_0 = const()[name = string("op_7037_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7037_end_mask_0 = const()[name = string("op_7037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7037_cast_fp16 = slice_by_index(begin = var_7037_begin_0, end = var_7037_end_0, end_mask = var_7037_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7037_cast_fp16")];
            tensor<int32, [4]> var_7042_begin_0 = const()[name = string("op_7042_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7042_end_0 = const()[name = string("op_7042_end_0"), val = tensor<int32, [4]>([19, 8, 2064, 128])];
            tensor<bool, [4]> var_7042_end_mask_0 = const()[name = string("op_7042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7042_cast_fp16 = slice_by_index(begin = var_7042_begin_0, end = var_7042_end_0, end_mask = var_7042_end_mask_0, x = read_state_1)[name = string("op_7042_cast_fp16")];
            tensor<int32, [4]> var_7052_begin_0 = const()[name = string("op_7052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7052_end_0 = const()[name = string("op_7052_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7052_end_mask_0 = const()[name = string("op_7052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7052_cast_fp16 = slice_by_index(begin = var_7052_begin_0, end = var_7052_end_0, end_mask = var_7052_end_mask_0, x = var_7042_cast_fp16)[name = string("op_7052_cast_fp16")];
            bool var_7058_transpose_x_0 = const()[name = string("op_7058_transpose_x_0"), val = bool(false)];
            bool var_7058_transpose_y_0 = const()[name = string("op_7058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7058_cast_fp16 = matmul(transpose_x = var_7058_transpose_x_0, transpose_y = var_7058_transpose_y_0, x = var_7037_cast_fp16, y = var_7052_cast_fp16)[name = string("op_7058_cast_fp16")];
            tensor<int32, [4]> var_7063_begin_0 = const()[name = string("op_7063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7063_end_0 = const()[name = string("op_7063_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7063_end_mask_0 = const()[name = string("op_7063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7063_cast_fp16 = slice_by_index(begin = var_7063_begin_0, end = var_7063_end_0, end_mask = var_7063_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7063_cast_fp16")];
            bool var_7064_transpose_x_0 = const()[name = string("op_7064_transpose_x_0"), val = bool(false)];
            bool var_7064_transpose_y_0 = const()[name = string("op_7064_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_6919_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7064_cast_fp16 = matmul(transpose_x = var_7064_transpose_x_0, transpose_y = var_7064_transpose_y_0, x = var_7063_cast_fp16, y = v_37_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7058_cast_fp16, y = var_7064_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7081_begin_0 = const()[name = string("op_7081_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7081_end_0 = const()[name = string("op_7081_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7081_end_mask_0 = const()[name = string("op_7081_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7081_cast_fp16 = slice_by_index(begin = var_7081_begin_0, end = var_7081_end_0, end_mask = var_7081_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7081_cast_fp16")];
            tensor<int32, [4]> var_7096_begin_0 = const()[name = string("op_7096_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7096_end_0 = const()[name = string("op_7096_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7096_end_mask_0 = const()[name = string("op_7096_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7096_cast_fp16 = slice_by_index(begin = var_7096_begin_0, end = var_7096_end_0, end_mask = var_7096_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7096_cast_fp16")];
            int32 var_7098 = const()[name = string("op_7098"), val = int32(1)];
            bool mixed_75_interleave_0 = const()[name = string("mixed_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_75_cast_fp16 = concat(axis = var_7098, interleave = mixed_75_interleave_0, values = (var_7081_cast_fp16, var_7096_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7102_perm_0 = const()[name = string("op_7102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7106 = const()[name = string("op_7106"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7102_cast_fp16 = transpose(perm = var_7102_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7106, x = var_7102_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599115840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601213056))))[name = string("model_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = linear_129_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_7117_promoted_to_fp16 = const()[name = string("op_7117_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7119_cast_fp16 = pow(x = x_227_cast_fp16, y = var_7117_promoted_to_fp16)[name = string("op_7119_cast_fp16")];
            tensor<int32, [1]> var_7121_axes_0 = const()[name = string("op_7121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7121_keep_dims_0 = const()[name = string("op_7121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7121_cast_fp16 = reduce_mean(axes = var_7121_axes_0, keep_dims = var_7121_keep_dims_0, x = var_7119_cast_fp16)[name = string("op_7121_cast_fp16")];
            fp16 var_7122_to_fp16 = const()[name = string("op_7122_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7123_cast_fp16 = add(x = var_7121_cast_fp16, y = var_7122_to_fp16)[name = string("op_7123_cast_fp16")];
            fp32 var_7124_epsilon_0 = const()[name = string("op_7124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7124_cast_fp16 = rsqrt(epsilon = var_7124_epsilon_0, x = var_7123_cast_fp16)[name = string("op_7124_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7125_cast_fp16 = mul(x = x_227_cast_fp16, y = var_7124_cast_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601215168)))];
            tensor<fp16, [1, 16, 1024]> input_185_cast_fp16 = mul(x = var_7125_cast_fp16, y = model_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601217280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604363072))))[name = string("model_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7130_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7130_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604369280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607515072))))[name = string("model_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_189_cast_fp16 = mul(x = var_7130_cast_fp16, y = linear_131_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607521280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610667072))))[name = string("model_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_132_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_229_cast_fp16 = add(x = x_227_cast_fp16, y = linear_132_cast_fp16)[name = string("x_229_cast_fp16")];
            fp16 var_7144_promoted_to_fp16 = const()[name = string("op_7144_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7146_cast_fp16 = pow(x = x_229_cast_fp16, y = var_7144_promoted_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<int32, [1]> var_7148_axes_0 = const()[name = string("op_7148_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7148_keep_dims_0 = const()[name = string("op_7148_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7148_cast_fp16 = reduce_mean(axes = var_7148_axes_0, keep_dims = var_7148_keep_dims_0, x = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp16 var_7149_to_fp16 = const()[name = string("op_7149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7150_cast_fp16 = add(x = var_7148_cast_fp16, y = var_7149_to_fp16)[name = string("op_7150_cast_fp16")];
            fp32 var_7151_epsilon_0 = const()[name = string("op_7151_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7151_cast_fp16 = rsqrt(epsilon = var_7151_epsilon_0, x = var_7150_cast_fp16)[name = string("op_7151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7152_cast_fp16 = mul(x = x_229_cast_fp16, y = var_7151_cast_fp16)[name = string("op_7152_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610669184)))];
            tensor<fp16, [1, 16, 1024]> input_191_cast_fp16 = mul(x = var_7152_cast_fp16, y = model_layers_19_input_layernorm_weight_to_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610671296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612768512))))[name = string("model_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_7161 = const()[name = string("op_7161"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7162_cast_fp16 = reshape(shape = var_7161, x = linear_133_cast_fp16)[name = string("op_7162_cast_fp16")];
            tensor<int32, [4]> x_231_perm_0 = const()[name = string("x_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7170_promoted_to_fp16 = const()[name = string("op_7170_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_231_cast_fp16 = transpose(perm = x_231_perm_0, x = var_7162_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7172_cast_fp16 = pow(x = x_231_cast_fp16, y = var_7170_promoted_to_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [1]> var_7174_axes_0 = const()[name = string("op_7174_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7174_keep_dims_0 = const()[name = string("op_7174_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7174_cast_fp16 = reduce_mean(axes = var_7174_axes_0, keep_dims = var_7174_keep_dims_0, x = var_7172_cast_fp16)[name = string("op_7174_cast_fp16")];
            fp16 var_7175_to_fp16 = const()[name = string("op_7175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7176_cast_fp16 = add(x = var_7174_cast_fp16, y = var_7175_to_fp16)[name = string("op_7176_cast_fp16")];
            fp32 var_7177_epsilon_0 = const()[name = string("op_7177_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7177_cast_fp16 = rsqrt(epsilon = var_7177_epsilon_0, x = var_7176_cast_fp16)[name = string("op_7177_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7178_cast_fp16 = mul(x = x_231_cast_fp16, y = var_7177_cast_fp16)[name = string("op_7178_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772672)))];
            tensor<fp16, [1, 16, 16, 128]> x_235_cast_fp16 = mul(x = var_7178_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613821632))))[name = string("model_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_7187 = const()[name = string("op_7187"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7188_cast_fp16 = reshape(shape = var_7187, x = linear_134_cast_fp16)[name = string("op_7188_cast_fp16")];
            tensor<int32, [4]> x_233_perm_0 = const()[name = string("x_233_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7196_promoted_to_fp16 = const()[name = string("op_7196_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_233_cast_fp16 = transpose(perm = x_233_perm_0, x = var_7188_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7198_cast_fp16 = pow(x = x_233_cast_fp16, y = var_7196_promoted_to_fp16)[name = string("op_7198_cast_fp16")];
            tensor<int32, [1]> var_7200_axes_0 = const()[name = string("op_7200_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7200_keep_dims_0 = const()[name = string("op_7200_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7200_cast_fp16 = reduce_mean(axes = var_7200_axes_0, keep_dims = var_7200_keep_dims_0, x = var_7198_cast_fp16)[name = string("op_7200_cast_fp16")];
            fp16 var_7201_to_fp16 = const()[name = string("op_7201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7202_cast_fp16 = add(x = var_7200_cast_fp16, y = var_7201_to_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7203_cast_fp16 = rsqrt(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7204_cast_fp16 = mul(x = x_233_cast_fp16, y = var_7203_cast_fp16)[name = string("op_7204_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613823744)))];
            tensor<fp16, [1, 8, 16, 128]> x_237_cast_fp16 = mul(x = var_7204_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613824064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614872704))))[name = string("model_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_7213 = const()[name = string("op_7213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7214_cast_fp16 = reshape(shape = var_7213, x = linear_135_cast_fp16)[name = string("op_7214_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_77_begin_0 = const()[name = string("first_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_77_end_0 = const()[name = string("first_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_77_end_mask_0 = const()[name = string("first_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_77_cast_fp16 = slice_by_index(begin = first_77_begin_0, end = first_77_end_0, end_mask = first_77_end_mask_0, x = x_235_cast_fp16)[name = string("first_77_cast_fp16")];
            tensor<int32, [4]> second_77_begin_0 = const()[name = string("second_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_77_end_0 = const()[name = string("second_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_77_end_mask_0 = const()[name = string("second_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_77_cast_fp16 = slice_by_index(begin = second_77_begin_0, end = second_77_end_0, end_mask = second_77_end_mask_0, x = x_235_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7228_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7229_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7231_cast_fp16 = sub(x = var_7228_cast_fp16, y = var_7229_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7232_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7233_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7235_cast_fp16 = add(x = var_7232_cast_fp16, y = var_7233_cast_fp16)[name = string("op_7235_cast_fp16")];
            int32 var_7237 = const()[name = string("op_7237"), val = int32(-1)];
            bool q_39_interleave_0 = const()[name = string("q_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_39_cast_fp16 = concat(axis = var_7237, interleave = q_39_interleave_0, values = (var_7231_cast_fp16, var_7235_cast_fp16))[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> first_79_begin_0 = const()[name = string("first_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_79_end_0 = const()[name = string("first_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_79_end_mask_0 = const()[name = string("first_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_79_cast_fp16 = slice_by_index(begin = first_79_begin_0, end = first_79_end_0, end_mask = first_79_end_mask_0, x = x_237_cast_fp16)[name = string("first_79_cast_fp16")];
            tensor<int32, [4]> second_79_begin_0 = const()[name = string("second_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_79_end_0 = const()[name = string("second_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_79_end_mask_0 = const()[name = string("second_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_79_cast_fp16 = slice_by_index(begin = second_79_begin_0, end = second_79_end_0, end_mask = second_79_end_mask_0, x = x_237_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7249_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7250_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7252_cast_fp16 = sub(x = var_7249_cast_fp16, y = var_7250_cast_fp16)[name = string("op_7252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7253_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7254_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7256_cast_fp16 = add(x = var_7253_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7256_cast_fp16")];
            int32 var_7258 = const()[name = string("op_7258"), val = int32(-1)];
            bool k_39_interleave_0 = const()[name = string("k_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_39_cast_fp16 = concat(axis = var_7258, interleave = k_39_interleave_0, values = (var_7252_cast_fp16, var_7256_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7269_begin_0 = const()[name = string("op_7269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7269_end_0 = const()[name = string("op_7269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7269_end_mask_0 = const()[name = string("op_7269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7269_stride_0 = const()[name = string("op_7269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7269_cast_fp16 = slice_by_index(begin = var_7269_begin_0, end = var_7269_end_0, end_mask = var_7269_end_mask_0, stride = var_7269_stride_0, x = q_39_cast_fp16)[name = string("op_7269_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7279_stride_0 = const()[name = string("op_7279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, stride = var_7279_stride_0, x = q_39_cast_fp16)[name = string("op_7279_cast_fp16")];
            int32 var_7281 = const()[name = string("op_7281"), val = int32(2)];
            bool stacked_39_interleave_0 = const()[name = string("stacked_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_39_cast_fp16 = concat(axis = var_7281, interleave = stacked_39_interleave_0, values = (var_7269_cast_fp16, var_7279_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7287_begin_0 = const()[name = string("op_7287_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7287_end_0 = const()[name = string("op_7287_end_0"), val = tensor<int32, [4]>([20, 8, 2064, 128])];
            tensor<bool, [4]> var_7287_end_mask_0 = const()[name = string("op_7287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7287_cast_fp16 = slice_by_index(begin = var_7287_begin_0, end = var_7287_end_0, end_mask = var_7287_end_mask_0, x = read_state_0)[name = string("op_7287_cast_fp16")];
            tensor<int32, [4]> var_7297_begin_0 = const()[name = string("op_7297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7297_end_0 = const()[name = string("op_7297_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7297_end_mask_0 = const()[name = string("op_7297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7297_cast_fp16 = slice_by_index(begin = var_7297_begin_0, end = var_7297_end_0, end_mask = var_7297_end_mask_0, x = var_7287_cast_fp16)[name = string("op_7297_cast_fp16")];
            bool var_7306_transpose_x_1 = const()[name = string("op_7306_transpose_x_1"), val = bool(false)];
            bool var_7306_transpose_y_1 = const()[name = string("op_7306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7306_cast_fp16 = matmul(transpose_x = var_7306_transpose_x_1, transpose_y = var_7306_transpose_y_1, x = var_7297_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7306_cast_fp16")];
            tensor<int32, [4]> var_7309_perm_0 = const()[name = string("op_7309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7310_to_fp16 = const()[name = string("op_7310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7309_cast_fp16 = transpose(perm = var_7309_perm_0, x = var_7306_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 2048]> var_7311_cast_fp16 = mul(x = var_7309_cast_fp16, y = var_7310_to_fp16)[name = string("op_7311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_39_cast_fp16 = add(x = var_7311_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7317_transpose_x_1 = const()[name = string("op_7317_transpose_x_1"), val = bool(false)];
            bool var_7317_transpose_y_1 = const()[name = string("op_7317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7317_cast_fp16 = matmul(transpose_x = var_7317_transpose_x_1, transpose_y = var_7317_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7317_cast_fp16")];
            fp16 var_7318_to_fp16 = const()[name = string("op_7318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7319_cast_fp16 = mul(x = var_7317_cast_fp16, y = var_7318_to_fp16)[name = string("op_7319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7323 = const()[name = string("op_7323"), val = int32(-1)];
            bool var_7324_interleave_0 = const()[name = string("op_7324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7324_cast_fp16 = concat(axis = var_7323, interleave = var_7324_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7324_cast_fp16")];
            int32 var_7325 = const()[name = string("op_7325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_39_cast_fp16 = softmax(axis = var_7325, x = var_7324_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7332_begin_0 = const()[name = string("op_7332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7332_end_0 = const()[name = string("op_7332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7332_end_mask_0 = const()[name = string("op_7332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7332_cast_fp16 = slice_by_index(begin = var_7332_begin_0, end = var_7332_end_0, end_mask = var_7332_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7332_cast_fp16")];
            tensor<int32, [4]> var_7337_begin_0 = const()[name = string("op_7337_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7337_end_0 = const()[name = string("op_7337_end_0"), val = tensor<int32, [4]>([20, 8, 2064, 128])];
            tensor<bool, [4]> var_7337_end_mask_0 = const()[name = string("op_7337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7337_cast_fp16 = slice_by_index(begin = var_7337_begin_0, end = var_7337_end_0, end_mask = var_7337_end_mask_0, x = read_state_1)[name = string("op_7337_cast_fp16")];
            tensor<int32, [4]> var_7347_begin_0 = const()[name = string("op_7347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7347_end_0 = const()[name = string("op_7347_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7347_end_mask_0 = const()[name = string("op_7347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7347_cast_fp16 = slice_by_index(begin = var_7347_begin_0, end = var_7347_end_0, end_mask = var_7347_end_mask_0, x = var_7337_cast_fp16)[name = string("op_7347_cast_fp16")];
            bool var_7353_transpose_x_0 = const()[name = string("op_7353_transpose_x_0"), val = bool(false)];
            bool var_7353_transpose_y_0 = const()[name = string("op_7353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7353_cast_fp16 = matmul(transpose_x = var_7353_transpose_x_0, transpose_y = var_7353_transpose_y_0, x = var_7332_cast_fp16, y = var_7347_cast_fp16)[name = string("op_7353_cast_fp16")];
            tensor<int32, [4]> var_7358_begin_0 = const()[name = string("op_7358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7358_end_0 = const()[name = string("op_7358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7358_end_mask_0 = const()[name = string("op_7358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7358_cast_fp16 = slice_by_index(begin = var_7358_begin_0, end = var_7358_end_0, end_mask = var_7358_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7358_cast_fp16")];
            bool var_7359_transpose_x_0 = const()[name = string("op_7359_transpose_x_0"), val = bool(false)];
            bool var_7359_transpose_y_0 = const()[name = string("op_7359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7214_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7359_cast_fp16 = matmul(transpose_x = var_7359_transpose_x_0, transpose_y = var_7359_transpose_y_0, x = var_7358_cast_fp16, y = v_39_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7353_cast_fp16, y = var_7359_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7376_begin_0 = const()[name = string("op_7376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7376_end_0 = const()[name = string("op_7376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7376_end_mask_0 = const()[name = string("op_7376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7376_cast_fp16 = slice_by_index(begin = var_7376_begin_0, end = var_7376_end_0, end_mask = var_7376_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7376_cast_fp16")];
            tensor<int32, [4]> var_7391_begin_0 = const()[name = string("op_7391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7391_end_0 = const()[name = string("op_7391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7391_end_mask_0 = const()[name = string("op_7391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7391_cast_fp16")];
            int32 var_7393 = const()[name = string("op_7393"), val = int32(1)];
            bool mixed_79_interleave_0 = const()[name = string("mixed_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_79_cast_fp16 = concat(axis = var_7393, interleave = mixed_79_interleave_0, values = (var_7376_cast_fp16, var_7391_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7397_perm_0 = const()[name = string("op_7397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7401 = const()[name = string("op_7401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7397_cast_fp16 = transpose(perm = var_7397_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7401, x = var_7397_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614874816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616972032))))[name = string("model_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = add(x = x_229_cast_fp16, y = linear_136_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_7412_promoted_to_fp16 = const()[name = string("op_7412_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7414_cast_fp16 = pow(x = x_239_cast_fp16, y = var_7412_promoted_to_fp16)[name = string("op_7414_cast_fp16")];
            tensor<int32, [1]> var_7416_axes_0 = const()[name = string("op_7416_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7416_keep_dims_0 = const()[name = string("op_7416_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7416_cast_fp16 = reduce_mean(axes = var_7416_axes_0, keep_dims = var_7416_keep_dims_0, x = var_7414_cast_fp16)[name = string("op_7416_cast_fp16")];
            fp16 var_7417_to_fp16 = const()[name = string("op_7417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7418_cast_fp16 = add(x = var_7416_cast_fp16, y = var_7417_to_fp16)[name = string("op_7418_cast_fp16")];
            fp32 var_7419_epsilon_0 = const()[name = string("op_7419_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7419_cast_fp16 = rsqrt(epsilon = var_7419_epsilon_0, x = var_7418_cast_fp16)[name = string("op_7419_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7420_cast_fp16 = mul(x = x_239_cast_fp16, y = var_7419_cast_fp16)[name = string("op_7420_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616974144)))];
            tensor<fp16, [1, 16, 1024]> input_195_cast_fp16 = mul(x = var_7420_cast_fp16, y = model_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616976256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620122048))))[name = string("model_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7425_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7425_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620128256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623274048))))[name = string("model_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_199_cast_fp16 = mul(x = var_7425_cast_fp16, y = linear_138_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623280256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626426048))))[name = string("model_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_241_cast_fp16 = add(x = x_239_cast_fp16, y = linear_139_cast_fp16)[name = string("x_241_cast_fp16")];
            fp16 var_7439_promoted_to_fp16 = const()[name = string("op_7439_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7441_cast_fp16 = pow(x = x_241_cast_fp16, y = var_7439_promoted_to_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7443_axes_0 = const()[name = string("op_7443_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7443_keep_dims_0 = const()[name = string("op_7443_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7443_cast_fp16 = reduce_mean(axes = var_7443_axes_0, keep_dims = var_7443_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7443_cast_fp16")];
            fp16 var_7444_to_fp16 = const()[name = string("op_7444_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7445_cast_fp16 = add(x = var_7443_cast_fp16, y = var_7444_to_fp16)[name = string("op_7445_cast_fp16")];
            fp32 var_7446_epsilon_0 = const()[name = string("op_7446_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7446_cast_fp16 = rsqrt(epsilon = var_7446_epsilon_0, x = var_7445_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7447_cast_fp16 = mul(x = x_241_cast_fp16, y = var_7446_cast_fp16)[name = string("op_7447_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626428160)))];
            tensor<fp16, [1, 16, 1024]> input_201_cast_fp16 = mul(x = var_7447_cast_fp16, y = model_layers_20_input_layernorm_weight_to_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626430272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628527488))))[name = string("model_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_7456 = const()[name = string("op_7456"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7457_cast_fp16 = reshape(shape = var_7456, x = linear_140_cast_fp16)[name = string("op_7457_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7465_promoted_to_fp16 = const()[name = string("op_7465_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_7457_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_7467_cast_fp16 = pow(x = x_243_cast_fp16, y = var_7465_promoted_to_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7469_axes_0 = const()[name = string("op_7469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7469_keep_dims_0 = const()[name = string("op_7469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7469_cast_fp16 = reduce_mean(axes = var_7469_axes_0, keep_dims = var_7469_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7469_cast_fp16")];
            fp16 var_7470_to_fp16 = const()[name = string("op_7470_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7471_cast_fp16 = add(x = var_7469_cast_fp16, y = var_7470_to_fp16)[name = string("op_7471_cast_fp16")];
            fp32 var_7472_epsilon_0 = const()[name = string("op_7472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7472_cast_fp16 = rsqrt(epsilon = var_7472_epsilon_0, x = var_7471_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7473_cast_fp16 = mul(x = x_243_cast_fp16, y = var_7472_cast_fp16)[name = string("op_7473_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531648)))];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = mul(x = var_7473_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_247_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629580608))))[name = string("model_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_7482 = const()[name = string("op_7482"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7483_cast_fp16 = reshape(shape = var_7482, x = linear_141_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7491_promoted_to_fp16 = const()[name = string("op_7491_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_7483_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_7493_cast_fp16 = pow(x = x_245_cast_fp16, y = var_7491_promoted_to_fp16)[name = string("op_7493_cast_fp16")];
            tensor<int32, [1]> var_7495_axes_0 = const()[name = string("op_7495_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7495_keep_dims_0 = const()[name = string("op_7495_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7495_cast_fp16 = reduce_mean(axes = var_7495_axes_0, keep_dims = var_7495_keep_dims_0, x = var_7493_cast_fp16)[name = string("op_7495_cast_fp16")];
            fp16 var_7496_to_fp16 = const()[name = string("op_7496_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7497_cast_fp16 = add(x = var_7495_cast_fp16, y = var_7496_to_fp16)[name = string("op_7497_cast_fp16")];
            fp32 var_7498_epsilon_0 = const()[name = string("op_7498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7498_cast_fp16 = rsqrt(epsilon = var_7498_epsilon_0, x = var_7497_cast_fp16)[name = string("op_7498_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = mul(x = x_245_cast_fp16, y = var_7498_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629582720)))];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = mul(x = var_7499_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629583040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630631680))))[name = string("model_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_7508 = const()[name = string("op_7508"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7509_cast_fp16 = reshape(shape = var_7508, x = linear_142_cast_fp16)[name = string("op_7509_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_81_begin_0 = const()[name = string("first_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_81_end_0 = const()[name = string("first_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_81_end_mask_0 = const()[name = string("first_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_81_cast_fp16 = slice_by_index(begin = first_81_begin_0, end = first_81_end_0, end_mask = first_81_end_mask_0, x = x_247_cast_fp16)[name = string("first_81_cast_fp16")];
            tensor<int32, [4]> second_81_begin_0 = const()[name = string("second_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_81_end_0 = const()[name = string("second_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_81_end_mask_0 = const()[name = string("second_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_81_cast_fp16 = slice_by_index(begin = second_81_begin_0, end = second_81_end_0, end_mask = second_81_end_mask_0, x = x_247_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7523_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7524_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7526_cast_fp16 = sub(x = var_7523_cast_fp16, y = var_7524_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7527_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7527_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7528_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7528_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7530_cast_fp16 = add(x = var_7527_cast_fp16, y = var_7528_cast_fp16)[name = string("op_7530_cast_fp16")];
            int32 var_7532 = const()[name = string("op_7532"), val = int32(-1)];
            bool q_41_interleave_0 = const()[name = string("q_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_41_cast_fp16 = concat(axis = var_7532, interleave = q_41_interleave_0, values = (var_7526_cast_fp16, var_7530_cast_fp16))[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> first_83_begin_0 = const()[name = string("first_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_83_end_0 = const()[name = string("first_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_83_end_mask_0 = const()[name = string("first_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_83_cast_fp16 = slice_by_index(begin = first_83_begin_0, end = first_83_end_0, end_mask = first_83_end_mask_0, x = x_249_cast_fp16)[name = string("first_83_cast_fp16")];
            tensor<int32, [4]> second_83_begin_0 = const()[name = string("second_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_83_end_0 = const()[name = string("second_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_83_end_mask_0 = const()[name = string("second_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_83_cast_fp16 = slice_by_index(begin = second_83_begin_0, end = second_83_end_0, end_mask = second_83_end_mask_0, x = x_249_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7544_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7545_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7547_cast_fp16 = sub(x = var_7544_cast_fp16, y = var_7545_cast_fp16)[name = string("op_7547_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7548_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7549_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7551_cast_fp16 = add(x = var_7548_cast_fp16, y = var_7549_cast_fp16)[name = string("op_7551_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool k_41_interleave_0 = const()[name = string("k_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_41_cast_fp16 = concat(axis = var_7553, interleave = k_41_interleave_0, values = (var_7547_cast_fp16, var_7551_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_7564_begin_0 = const()[name = string("op_7564_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7564_end_0 = const()[name = string("op_7564_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7564_end_mask_0 = const()[name = string("op_7564_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7564_stride_0 = const()[name = string("op_7564_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7564_cast_fp16 = slice_by_index(begin = var_7564_begin_0, end = var_7564_end_0, end_mask = var_7564_end_mask_0, stride = var_7564_stride_0, x = q_41_cast_fp16)[name = string("op_7564_cast_fp16")];
            tensor<int32, [4]> var_7574_begin_0 = const()[name = string("op_7574_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7574_end_0 = const()[name = string("op_7574_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7574_end_mask_0 = const()[name = string("op_7574_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7574_stride_0 = const()[name = string("op_7574_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7574_cast_fp16 = slice_by_index(begin = var_7574_begin_0, end = var_7574_end_0, end_mask = var_7574_end_mask_0, stride = var_7574_stride_0, x = q_41_cast_fp16)[name = string("op_7574_cast_fp16")];
            int32 var_7576 = const()[name = string("op_7576"), val = int32(2)];
            bool stacked_41_interleave_0 = const()[name = string("stacked_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_41_cast_fp16 = concat(axis = var_7576, interleave = stacked_41_interleave_0, values = (var_7564_cast_fp16, var_7574_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_7582_begin_0 = const()[name = string("op_7582_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7582_end_0 = const()[name = string("op_7582_end_0"), val = tensor<int32, [4]>([21, 8, 2064, 128])];
            tensor<bool, [4]> var_7582_end_mask_0 = const()[name = string("op_7582_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7582_cast_fp16 = slice_by_index(begin = var_7582_begin_0, end = var_7582_end_0, end_mask = var_7582_end_mask_0, x = read_state_0)[name = string("op_7582_cast_fp16")];
            tensor<int32, [4]> var_7592_begin_0 = const()[name = string("op_7592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7592_end_0 = const()[name = string("op_7592_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7592_end_mask_0 = const()[name = string("op_7592_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7592_cast_fp16 = slice_by_index(begin = var_7592_begin_0, end = var_7592_end_0, end_mask = var_7592_end_mask_0, x = var_7582_cast_fp16)[name = string("op_7592_cast_fp16")];
            bool var_7601_transpose_x_1 = const()[name = string("op_7601_transpose_x_1"), val = bool(false)];
            bool var_7601_transpose_y_1 = const()[name = string("op_7601_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7601_cast_fp16 = matmul(transpose_x = var_7601_transpose_x_1, transpose_y = var_7601_transpose_y_1, x = var_7592_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_7601_cast_fp16")];
            tensor<int32, [4]> var_7604_perm_0 = const()[name = string("op_7604_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7605_to_fp16 = const()[name = string("op_7605_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7604_cast_fp16 = transpose(perm = var_7604_perm_0, x = var_7601_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 2048]> var_7606_cast_fp16 = mul(x = var_7604_cast_fp16, y = var_7605_to_fp16)[name = string("op_7606_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_41_cast_fp16 = add(x = var_7606_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_7612_transpose_x_1 = const()[name = string("op_7612_transpose_x_1"), val = bool(false)];
            bool var_7612_transpose_y_1 = const()[name = string("op_7612_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7612_cast_fp16 = matmul(transpose_x = var_7612_transpose_x_1, transpose_y = var_7612_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_7612_cast_fp16")];
            fp16 var_7613_to_fp16 = const()[name = string("op_7613_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7614_cast_fp16 = mul(x = var_7612_cast_fp16, y = var_7613_to_fp16)[name = string("op_7614_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_7614_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_7618 = const()[name = string("op_7618"), val = int32(-1)];
            bool var_7619_interleave_0 = const()[name = string("op_7619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7619_cast_fp16 = concat(axis = var_7618, interleave = var_7619_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_7619_cast_fp16")];
            int32 var_7620 = const()[name = string("op_7620"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_41_cast_fp16 = softmax(axis = var_7620, x = var_7619_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_7627_begin_0 = const()[name = string("op_7627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7627_end_0 = const()[name = string("op_7627_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7627_end_mask_0 = const()[name = string("op_7627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7627_cast_fp16 = slice_by_index(begin = var_7627_begin_0, end = var_7627_end_0, end_mask = var_7627_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7627_cast_fp16")];
            tensor<int32, [4]> var_7632_begin_0 = const()[name = string("op_7632_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7632_end_0 = const()[name = string("op_7632_end_0"), val = tensor<int32, [4]>([21, 8, 2064, 128])];
            tensor<bool, [4]> var_7632_end_mask_0 = const()[name = string("op_7632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7632_cast_fp16 = slice_by_index(begin = var_7632_begin_0, end = var_7632_end_0, end_mask = var_7632_end_mask_0, x = read_state_1)[name = string("op_7632_cast_fp16")];
            tensor<int32, [4]> var_7642_begin_0 = const()[name = string("op_7642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7642_end_0 = const()[name = string("op_7642_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7642_end_mask_0 = const()[name = string("op_7642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7642_cast_fp16 = slice_by_index(begin = var_7642_begin_0, end = var_7642_end_0, end_mask = var_7642_end_mask_0, x = var_7632_cast_fp16)[name = string("op_7642_cast_fp16")];
            bool var_7648_transpose_x_0 = const()[name = string("op_7648_transpose_x_0"), val = bool(false)];
            bool var_7648_transpose_y_0 = const()[name = string("op_7648_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7648_cast_fp16 = matmul(transpose_x = var_7648_transpose_x_0, transpose_y = var_7648_transpose_y_0, x = var_7627_cast_fp16, y = var_7642_cast_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [4]> var_7653_begin_0 = const()[name = string("op_7653_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7653_end_0 = const()[name = string("op_7653_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7653_end_mask_0 = const()[name = string("op_7653_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7653_cast_fp16 = slice_by_index(begin = var_7653_begin_0, end = var_7653_end_0, end_mask = var_7653_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7653_cast_fp16")];
            bool var_7654_transpose_x_0 = const()[name = string("op_7654_transpose_x_0"), val = bool(false)];
            bool var_7654_transpose_y_0 = const()[name = string("op_7654_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_7509_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_7654_cast_fp16 = matmul(transpose_x = var_7654_transpose_x_0, transpose_y = var_7654_transpose_y_0, x = var_7653_cast_fp16, y = v_41_cast_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_7648_cast_fp16, y = var_7654_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_7671_begin_0 = const()[name = string("op_7671_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7671_end_0 = const()[name = string("op_7671_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7671_end_mask_0 = const()[name = string("op_7671_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7671_cast_fp16 = slice_by_index(begin = var_7671_begin_0, end = var_7671_end_0, end_mask = var_7671_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<int32, [4]> var_7686_begin_0 = const()[name = string("op_7686_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7686_end_0 = const()[name = string("op_7686_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7686_end_mask_0 = const()[name = string("op_7686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7686_cast_fp16 = slice_by_index(begin = var_7686_begin_0, end = var_7686_end_0, end_mask = var_7686_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7686_cast_fp16")];
            int32 var_7688 = const()[name = string("op_7688"), val = int32(1)];
            bool mixed_83_interleave_0 = const()[name = string("mixed_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_83_cast_fp16 = concat(axis = var_7688, interleave = mixed_83_interleave_0, values = (var_7671_cast_fp16, var_7686_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_7692_perm_0 = const()[name = string("op_7692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7696 = const()[name = string("op_7696"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7692_cast_fp16 = transpose(perm = var_7692_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_7696, x = var_7692_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630633792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632731008))))[name = string("model_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_251_cast_fp16 = add(x = x_241_cast_fp16, y = linear_143_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7709_cast_fp16 = pow(x = x_251_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7709_cast_fp16")];
            tensor<int32, [1]> var_7711_axes_0 = const()[name = string("op_7711_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7711_keep_dims_0 = const()[name = string("op_7711_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7711_cast_fp16 = reduce_mean(axes = var_7711_axes_0, keep_dims = var_7711_keep_dims_0, x = var_7709_cast_fp16)[name = string("op_7711_cast_fp16")];
            fp16 var_7712_to_fp16 = const()[name = string("op_7712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7713_cast_fp16 = add(x = var_7711_cast_fp16, y = var_7712_to_fp16)[name = string("op_7713_cast_fp16")];
            fp32 var_7714_epsilon_0 = const()[name = string("op_7714_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7714_cast_fp16 = rsqrt(epsilon = var_7714_epsilon_0, x = var_7713_cast_fp16)[name = string("op_7714_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7715_cast_fp16 = mul(x = x_251_cast_fp16, y = var_7714_cast_fp16)[name = string("op_7715_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632733120)))];
            tensor<fp16, [1, 16, 1024]> input_205_cast_fp16 = mul(x = var_7715_cast_fp16, y = model_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632735232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635881024))))[name = string("model_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7720_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_7720_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635887232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639033024))))[name = string("model_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_209_cast_fp16 = mul(x = var_7720_cast_fp16, y = linear_145_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639039232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642185024))))[name = string("model_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_146_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_253_cast_fp16 = add(x = x_251_cast_fp16, y = linear_146_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_7734_promoted_to_fp16 = const()[name = string("op_7734_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7736_cast_fp16 = pow(x = x_253_cast_fp16, y = var_7734_promoted_to_fp16)[name = string("op_7736_cast_fp16")];
            tensor<int32, [1]> var_7738_axes_0 = const()[name = string("op_7738_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7738_keep_dims_0 = const()[name = string("op_7738_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7738_cast_fp16 = reduce_mean(axes = var_7738_axes_0, keep_dims = var_7738_keep_dims_0, x = var_7736_cast_fp16)[name = string("op_7738_cast_fp16")];
            fp16 var_7739_to_fp16 = const()[name = string("op_7739_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7740_cast_fp16 = add(x = var_7738_cast_fp16, y = var_7739_to_fp16)[name = string("op_7740_cast_fp16")];
            fp32 var_7741_epsilon_0 = const()[name = string("op_7741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7741_cast_fp16 = rsqrt(epsilon = var_7741_epsilon_0, x = var_7740_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7742_cast_fp16 = mul(x = x_253_cast_fp16, y = var_7741_cast_fp16)[name = string("op_7742_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642187136)))];
            tensor<fp16, [1, 16, 1024]> input_211_cast_fp16 = mul(x = var_7742_cast_fp16, y = model_layers_21_input_layernorm_weight_to_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642189248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644286464))))[name = string("model_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7751 = const()[name = string("op_7751"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7752_cast_fp16 = reshape(shape = var_7751, x = linear_147_cast_fp16)[name = string("op_7752_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7760_promoted_to_fp16 = const()[name = string("op_7760_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_7752_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_7762_cast_fp16 = pow(x = x_255_cast_fp16, y = var_7760_promoted_to_fp16)[name = string("op_7762_cast_fp16")];
            tensor<int32, [1]> var_7764_axes_0 = const()[name = string("op_7764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7764_keep_dims_0 = const()[name = string("op_7764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7764_cast_fp16 = reduce_mean(axes = var_7764_axes_0, keep_dims = var_7764_keep_dims_0, x = var_7762_cast_fp16)[name = string("op_7764_cast_fp16")];
            fp16 var_7765_to_fp16 = const()[name = string("op_7765_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7766_cast_fp16 = add(x = var_7764_cast_fp16, y = var_7765_to_fp16)[name = string("op_7766_cast_fp16")];
            fp32 var_7767_epsilon_0 = const()[name = string("op_7767_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7767_cast_fp16 = rsqrt(epsilon = var_7767_epsilon_0, x = var_7766_cast_fp16)[name = string("op_7767_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7768_cast_fp16 = mul(x = x_255_cast_fp16, y = var_7767_cast_fp16)[name = string("op_7768_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290624)))];
            tensor<fp16, [1, 16, 16, 128]> x_259_cast_fp16 = mul(x = var_7768_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645339584))))[name = string("model_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_7777 = const()[name = string("op_7777"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7778_cast_fp16 = reshape(shape = var_7777, x = linear_148_cast_fp16)[name = string("op_7778_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7786_promoted_to_fp16 = const()[name = string("op_7786_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_7778_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_7788_cast_fp16 = pow(x = x_257_cast_fp16, y = var_7786_promoted_to_fp16)[name = string("op_7788_cast_fp16")];
            tensor<int32, [1]> var_7790_axes_0 = const()[name = string("op_7790_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7790_keep_dims_0 = const()[name = string("op_7790_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7790_cast_fp16 = reduce_mean(axes = var_7790_axes_0, keep_dims = var_7790_keep_dims_0, x = var_7788_cast_fp16)[name = string("op_7790_cast_fp16")];
            fp16 var_7791_to_fp16 = const()[name = string("op_7791_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7792_cast_fp16 = add(x = var_7790_cast_fp16, y = var_7791_to_fp16)[name = string("op_7792_cast_fp16")];
            fp32 var_7793_epsilon_0 = const()[name = string("op_7793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7793_cast_fp16 = rsqrt(epsilon = var_7793_epsilon_0, x = var_7792_cast_fp16)[name = string("op_7793_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7794_cast_fp16 = mul(x = x_257_cast_fp16, y = var_7793_cast_fp16)[name = string("op_7794_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645341696)))];
            tensor<fp16, [1, 8, 16, 128]> x_261_cast_fp16 = mul(x = var_7794_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645342016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646390656))))[name = string("model_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_7803 = const()[name = string("op_7803"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7804_cast_fp16 = reshape(shape = var_7803, x = linear_149_cast_fp16)[name = string("op_7804_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_85_begin_0 = const()[name = string("first_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_85_end_0 = const()[name = string("first_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_85_end_mask_0 = const()[name = string("first_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_85_cast_fp16 = slice_by_index(begin = first_85_begin_0, end = first_85_end_0, end_mask = first_85_end_mask_0, x = x_259_cast_fp16)[name = string("first_85_cast_fp16")];
            tensor<int32, [4]> second_85_begin_0 = const()[name = string("second_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_85_end_0 = const()[name = string("second_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_85_end_mask_0 = const()[name = string("second_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_85_cast_fp16 = slice_by_index(begin = second_85_begin_0, end = second_85_end_0, end_mask = second_85_end_mask_0, x = x_259_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7818_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7818_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7819_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7819_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7821_cast_fp16 = sub(x = var_7818_cast_fp16, y = var_7819_cast_fp16)[name = string("op_7821_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7822_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7823_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7823_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7825_cast_fp16 = add(x = var_7822_cast_fp16, y = var_7823_cast_fp16)[name = string("op_7825_cast_fp16")];
            int32 var_7827 = const()[name = string("op_7827"), val = int32(-1)];
            bool q_43_interleave_0 = const()[name = string("q_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_43_cast_fp16 = concat(axis = var_7827, interleave = q_43_interleave_0, values = (var_7821_cast_fp16, var_7825_cast_fp16))[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> first_87_begin_0 = const()[name = string("first_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_87_end_0 = const()[name = string("first_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_87_end_mask_0 = const()[name = string("first_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_87_cast_fp16 = slice_by_index(begin = first_87_begin_0, end = first_87_end_0, end_mask = first_87_end_mask_0, x = x_261_cast_fp16)[name = string("first_87_cast_fp16")];
            tensor<int32, [4]> second_87_begin_0 = const()[name = string("second_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_87_end_0 = const()[name = string("second_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_87_end_mask_0 = const()[name = string("second_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_87_cast_fp16 = slice_by_index(begin = second_87_begin_0, end = second_87_end_0, end_mask = second_87_end_mask_0, x = x_261_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7839_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7839_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7840_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7840_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7842_cast_fp16 = sub(x = var_7839_cast_fp16, y = var_7840_cast_fp16)[name = string("op_7842_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7843_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7844_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7846_cast_fp16 = add(x = var_7843_cast_fp16, y = var_7844_cast_fp16)[name = string("op_7846_cast_fp16")];
            int32 var_7848 = const()[name = string("op_7848"), val = int32(-1)];
            bool k_43_interleave_0 = const()[name = string("k_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_43_cast_fp16 = concat(axis = var_7848, interleave = k_43_interleave_0, values = (var_7842_cast_fp16, var_7846_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_7859_begin_0 = const()[name = string("op_7859_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7859_end_0 = const()[name = string("op_7859_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7859_end_mask_0 = const()[name = string("op_7859_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7859_stride_0 = const()[name = string("op_7859_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7859_cast_fp16 = slice_by_index(begin = var_7859_begin_0, end = var_7859_end_0, end_mask = var_7859_end_mask_0, stride = var_7859_stride_0, x = q_43_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7869_begin_0 = const()[name = string("op_7869_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7869_end_0 = const()[name = string("op_7869_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7869_end_mask_0 = const()[name = string("op_7869_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7869_stride_0 = const()[name = string("op_7869_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7869_cast_fp16 = slice_by_index(begin = var_7869_begin_0, end = var_7869_end_0, end_mask = var_7869_end_mask_0, stride = var_7869_stride_0, x = q_43_cast_fp16)[name = string("op_7869_cast_fp16")];
            int32 var_7871 = const()[name = string("op_7871"), val = int32(2)];
            bool stacked_43_interleave_0 = const()[name = string("stacked_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_43_cast_fp16 = concat(axis = var_7871, interleave = stacked_43_interleave_0, values = (var_7859_cast_fp16, var_7869_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_7877_begin_0 = const()[name = string("op_7877_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7877_end_0 = const()[name = string("op_7877_end_0"), val = tensor<int32, [4]>([22, 8, 2064, 128])];
            tensor<bool, [4]> var_7877_end_mask_0 = const()[name = string("op_7877_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7877_cast_fp16 = slice_by_index(begin = var_7877_begin_0, end = var_7877_end_0, end_mask = var_7877_end_mask_0, x = read_state_0)[name = string("op_7877_cast_fp16")];
            tensor<int32, [4]> var_7887_begin_0 = const()[name = string("op_7887_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7887_end_0 = const()[name = string("op_7887_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7887_end_mask_0 = const()[name = string("op_7887_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7887_cast_fp16 = slice_by_index(begin = var_7887_begin_0, end = var_7887_end_0, end_mask = var_7887_end_mask_0, x = var_7877_cast_fp16)[name = string("op_7887_cast_fp16")];
            bool var_7896_transpose_x_1 = const()[name = string("op_7896_transpose_x_1"), val = bool(false)];
            bool var_7896_transpose_y_1 = const()[name = string("op_7896_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_7896_cast_fp16 = matmul(transpose_x = var_7896_transpose_x_1, transpose_y = var_7896_transpose_y_1, x = var_7887_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_7896_cast_fp16")];
            tensor<int32, [4]> var_7899_perm_0 = const()[name = string("op_7899_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7900_to_fp16 = const()[name = string("op_7900_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_7899_cast_fp16 = transpose(perm = var_7899_perm_0, x = var_7896_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 2048]> var_7901_cast_fp16 = mul(x = var_7899_cast_fp16, y = var_7900_to_fp16)[name = string("op_7901_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_43_cast_fp16 = add(x = var_7901_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_7907_transpose_x_1 = const()[name = string("op_7907_transpose_x_1"), val = bool(false)];
            bool var_7907_transpose_y_1 = const()[name = string("op_7907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7907_cast_fp16 = matmul(transpose_x = var_7907_transpose_x_1, transpose_y = var_7907_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_7907_cast_fp16")];
            fp16 var_7908_to_fp16 = const()[name = string("op_7908_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7909_cast_fp16 = mul(x = var_7907_cast_fp16, y = var_7908_to_fp16)[name = string("op_7909_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_7909_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_7913 = const()[name = string("op_7913"), val = int32(-1)];
            bool var_7914_interleave_0 = const()[name = string("op_7914_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_7914_cast_fp16 = concat(axis = var_7913, interleave = var_7914_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_7914_cast_fp16")];
            int32 var_7915 = const()[name = string("op_7915"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_43_cast_fp16 = softmax(axis = var_7915, x = var_7914_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_7922_begin_0 = const()[name = string("op_7922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7922_end_0 = const()[name = string("op_7922_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_7922_end_mask_0 = const()[name = string("op_7922_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([22, 8, 2064, 128])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = read_state_1)[name = string("op_7927_cast_fp16")];
            tensor<int32, [4]> var_7937_begin_0 = const()[name = string("op_7937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7937_end_0 = const()[name = string("op_7937_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_7937_end_mask_0 = const()[name = string("op_7937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_7937_cast_fp16 = slice_by_index(begin = var_7937_begin_0, end = var_7937_end_0, end_mask = var_7937_end_mask_0, x = var_7927_cast_fp16)[name = string("op_7937_cast_fp16")];
            bool var_7943_transpose_x_0 = const()[name = string("op_7943_transpose_x_0"), val = bool(false)];
            bool var_7943_transpose_y_0 = const()[name = string("op_7943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7943_cast_fp16 = matmul(transpose_x = var_7943_transpose_x_0, transpose_y = var_7943_transpose_y_0, x = var_7922_cast_fp16, y = var_7937_cast_fp16)[name = string("op_7943_cast_fp16")];
            tensor<int32, [4]> var_7948_begin_0 = const()[name = string("op_7948_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_7948_end_0 = const()[name = string("op_7948_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_7948_end_mask_0 = const()[name = string("op_7948_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7948_cast_fp16 = slice_by_index(begin = var_7948_begin_0, end = var_7948_end_0, end_mask = var_7948_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7948_cast_fp16")];
            bool var_7949_transpose_x_0 = const()[name = string("op_7949_transpose_x_0"), val = bool(false)];
            bool var_7949_transpose_y_0 = const()[name = string("op_7949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_7804_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_7949_cast_fp16 = matmul(transpose_x = var_7949_transpose_x_0, transpose_y = var_7949_transpose_y_0, x = var_7948_cast_fp16, y = v_43_cast_fp16)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_7943_cast_fp16, y = var_7949_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_7966_begin_0 = const()[name = string("op_7966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7966_end_0 = const()[name = string("op_7966_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7966_end_mask_0 = const()[name = string("op_7966_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7966_cast_fp16 = slice_by_index(begin = var_7966_begin_0, end = var_7966_end_0, end_mask = var_7966_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7966_cast_fp16")];
            tensor<int32, [4]> var_7981_begin_0 = const()[name = string("op_7981_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7981_end_0 = const()[name = string("op_7981_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7981_end_mask_0 = const()[name = string("op_7981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7981_cast_fp16 = slice_by_index(begin = var_7981_begin_0, end = var_7981_end_0, end_mask = var_7981_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7981_cast_fp16")];
            int32 var_7983 = const()[name = string("op_7983"), val = int32(1)];
            bool mixed_87_interleave_0 = const()[name = string("mixed_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_87_cast_fp16 = concat(axis = var_7983, interleave = mixed_87_interleave_0, values = (var_7966_cast_fp16, var_7981_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_7987_perm_0 = const()[name = string("op_7987_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7991 = const()[name = string("op_7991"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7987_cast_fp16 = transpose(perm = var_7987_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_7991, x = var_7987_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646392768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648489984))))[name = string("model_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_150_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = linear_150_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_8002_promoted_to_fp16 = const()[name = string("op_8002_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8004_cast_fp16 = pow(x = x_263_cast_fp16, y = var_8002_promoted_to_fp16)[name = string("op_8004_cast_fp16")];
            tensor<int32, [1]> var_8006_axes_0 = const()[name = string("op_8006_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8006_keep_dims_0 = const()[name = string("op_8006_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8006_cast_fp16 = reduce_mean(axes = var_8006_axes_0, keep_dims = var_8006_keep_dims_0, x = var_8004_cast_fp16)[name = string("op_8006_cast_fp16")];
            fp16 var_8007_to_fp16 = const()[name = string("op_8007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = add(x = var_8006_cast_fp16, y = var_8007_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = rsqrt(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8010_cast_fp16 = mul(x = x_263_cast_fp16, y = var_8009_cast_fp16)[name = string("op_8010_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648492096)))];
            tensor<fp16, [1, 16, 1024]> input_215_cast_fp16 = mul(x = var_8010_cast_fp16, y = model_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648494208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651640000))))[name = string("model_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8015_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8015_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651646208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654792000))))[name = string("model_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_219_cast_fp16 = mul(x = var_8015_cast_fp16, y = linear_152_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654798208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657944000))))[name = string("model_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_265_cast_fp16 = add(x = x_263_cast_fp16, y = linear_153_cast_fp16)[name = string("x_265_cast_fp16")];
            fp16 var_8029_promoted_to_fp16 = const()[name = string("op_8029_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8031_cast_fp16 = pow(x = x_265_cast_fp16, y = var_8029_promoted_to_fp16)[name = string("op_8031_cast_fp16")];
            tensor<int32, [1]> var_8033_axes_0 = const()[name = string("op_8033_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8033_keep_dims_0 = const()[name = string("op_8033_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8033_cast_fp16 = reduce_mean(axes = var_8033_axes_0, keep_dims = var_8033_keep_dims_0, x = var_8031_cast_fp16)[name = string("op_8033_cast_fp16")];
            fp16 var_8034_to_fp16 = const()[name = string("op_8034_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8035_cast_fp16 = add(x = var_8033_cast_fp16, y = var_8034_to_fp16)[name = string("op_8035_cast_fp16")];
            fp32 var_8036_epsilon_0 = const()[name = string("op_8036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8036_cast_fp16 = rsqrt(epsilon = var_8036_epsilon_0, x = var_8035_cast_fp16)[name = string("op_8036_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8037_cast_fp16 = mul(x = x_265_cast_fp16, y = var_8036_cast_fp16)[name = string("op_8037_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657946112)))];
            tensor<fp16, [1, 16, 1024]> input_221_cast_fp16 = mul(x = var_8037_cast_fp16, y = model_layers_22_input_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657948224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660045440))))[name = string("model_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_8046 = const()[name = string("op_8046"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = reshape(shape = var_8046, x = linear_154_cast_fp16)[name = string("op_8047_cast_fp16")];
            tensor<int32, [4]> x_267_perm_0 = const()[name = string("x_267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8055_promoted_to_fp16 = const()[name = string("op_8055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = transpose(perm = x_267_perm_0, x = var_8047_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8057_cast_fp16 = pow(x = x_267_cast_fp16, y = var_8055_promoted_to_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [1]> var_8059_axes_0 = const()[name = string("op_8059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8059_keep_dims_0 = const()[name = string("op_8059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8059_cast_fp16 = reduce_mean(axes = var_8059_axes_0, keep_dims = var_8059_keep_dims_0, x = var_8057_cast_fp16)[name = string("op_8059_cast_fp16")];
            fp16 var_8060_to_fp16 = const()[name = string("op_8060_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8061_cast_fp16 = add(x = var_8059_cast_fp16, y = var_8060_to_fp16)[name = string("op_8061_cast_fp16")];
            fp32 var_8062_epsilon_0 = const()[name = string("op_8062_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8062_cast_fp16 = rsqrt(epsilon = var_8062_epsilon_0, x = var_8061_cast_fp16)[name = string("op_8062_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8063_cast_fp16 = mul(x = x_267_cast_fp16, y = var_8062_cast_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049600)))];
            tensor<fp16, [1, 16, 16, 128]> x_271_cast_fp16 = mul(x = var_8063_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661098560))))[name = string("model_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_8072 = const()[name = string("op_8072"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8073_cast_fp16 = reshape(shape = var_8072, x = linear_155_cast_fp16)[name = string("op_8073_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8081_promoted_to_fp16 = const()[name = string("op_8081_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_8073_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8083_cast_fp16 = pow(x = x_269_cast_fp16, y = var_8081_promoted_to_fp16)[name = string("op_8083_cast_fp16")];
            tensor<int32, [1]> var_8085_axes_0 = const()[name = string("op_8085_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8085_keep_dims_0 = const()[name = string("op_8085_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8085_cast_fp16 = reduce_mean(axes = var_8085_axes_0, keep_dims = var_8085_keep_dims_0, x = var_8083_cast_fp16)[name = string("op_8085_cast_fp16")];
            fp16 var_8086_to_fp16 = const()[name = string("op_8086_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8087_cast_fp16 = add(x = var_8085_cast_fp16, y = var_8086_to_fp16)[name = string("op_8087_cast_fp16")];
            fp32 var_8088_epsilon_0 = const()[name = string("op_8088_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8088_cast_fp16 = rsqrt(epsilon = var_8088_epsilon_0, x = var_8087_cast_fp16)[name = string("op_8088_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8089_cast_fp16 = mul(x = x_269_cast_fp16, y = var_8088_cast_fp16)[name = string("op_8089_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100672)))];
            tensor<fp16, [1, 8, 16, 128]> x_273_cast_fp16 = mul(x = var_8089_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662149632))))[name = string("model_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_8098 = const()[name = string("op_8098"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8099_cast_fp16 = reshape(shape = var_8098, x = linear_156_cast_fp16)[name = string("op_8099_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_89_begin_0 = const()[name = string("first_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_89_end_0 = const()[name = string("first_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_89_end_mask_0 = const()[name = string("first_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_89_cast_fp16 = slice_by_index(begin = first_89_begin_0, end = first_89_end_0, end_mask = first_89_end_mask_0, x = x_271_cast_fp16)[name = string("first_89_cast_fp16")];
            tensor<int32, [4]> second_89_begin_0 = const()[name = string("second_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_89_end_0 = const()[name = string("second_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_89_end_mask_0 = const()[name = string("second_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_89_cast_fp16 = slice_by_index(begin = second_89_begin_0, end = second_89_end_0, end_mask = second_89_end_mask_0, x = x_271_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8113_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8114_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8114_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8116_cast_fp16 = sub(x = var_8113_cast_fp16, y = var_8114_cast_fp16)[name = string("op_8116_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8117_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8117_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8118_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8118_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8120_cast_fp16 = add(x = var_8117_cast_fp16, y = var_8118_cast_fp16)[name = string("op_8120_cast_fp16")];
            int32 var_8122 = const()[name = string("op_8122"), val = int32(-1)];
            bool q_45_interleave_0 = const()[name = string("q_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_45_cast_fp16 = concat(axis = var_8122, interleave = q_45_interleave_0, values = (var_8116_cast_fp16, var_8120_cast_fp16))[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> first_91_begin_0 = const()[name = string("first_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_91_end_0 = const()[name = string("first_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_91_end_mask_0 = const()[name = string("first_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_91_cast_fp16 = slice_by_index(begin = first_91_begin_0, end = first_91_end_0, end_mask = first_91_end_mask_0, x = x_273_cast_fp16)[name = string("first_91_cast_fp16")];
            tensor<int32, [4]> second_91_begin_0 = const()[name = string("second_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_91_end_0 = const()[name = string("second_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_91_end_mask_0 = const()[name = string("second_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_91_cast_fp16 = slice_by_index(begin = second_91_begin_0, end = second_91_end_0, end_mask = second_91_end_mask_0, x = x_273_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8134_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8134_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8135_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8135_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8137_cast_fp16 = sub(x = var_8134_cast_fp16, y = var_8135_cast_fp16)[name = string("op_8137_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8138_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8139_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8141_cast_fp16 = add(x = var_8138_cast_fp16, y = var_8139_cast_fp16)[name = string("op_8141_cast_fp16")];
            int32 var_8143 = const()[name = string("op_8143"), val = int32(-1)];
            bool k_45_interleave_0 = const()[name = string("k_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_45_cast_fp16 = concat(axis = var_8143, interleave = k_45_interleave_0, values = (var_8137_cast_fp16, var_8141_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8154_begin_0 = const()[name = string("op_8154_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8154_end_0 = const()[name = string("op_8154_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8154_end_mask_0 = const()[name = string("op_8154_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8154_stride_0 = const()[name = string("op_8154_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8154_cast_fp16 = slice_by_index(begin = var_8154_begin_0, end = var_8154_end_0, end_mask = var_8154_end_mask_0, stride = var_8154_stride_0, x = q_45_cast_fp16)[name = string("op_8154_cast_fp16")];
            tensor<int32, [4]> var_8164_begin_0 = const()[name = string("op_8164_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8164_end_0 = const()[name = string("op_8164_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8164_end_mask_0 = const()[name = string("op_8164_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8164_stride_0 = const()[name = string("op_8164_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8164_cast_fp16 = slice_by_index(begin = var_8164_begin_0, end = var_8164_end_0, end_mask = var_8164_end_mask_0, stride = var_8164_stride_0, x = q_45_cast_fp16)[name = string("op_8164_cast_fp16")];
            int32 var_8166 = const()[name = string("op_8166"), val = int32(2)];
            bool stacked_45_interleave_0 = const()[name = string("stacked_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_45_cast_fp16 = concat(axis = var_8166, interleave = stacked_45_interleave_0, values = (var_8154_cast_fp16, var_8164_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8172_begin_0 = const()[name = string("op_8172_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8172_end_0 = const()[name = string("op_8172_end_0"), val = tensor<int32, [4]>([23, 8, 2064, 128])];
            tensor<bool, [4]> var_8172_end_mask_0 = const()[name = string("op_8172_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8172_cast_fp16 = slice_by_index(begin = var_8172_begin_0, end = var_8172_end_0, end_mask = var_8172_end_mask_0, x = read_state_0)[name = string("op_8172_cast_fp16")];
            tensor<int32, [4]> var_8182_begin_0 = const()[name = string("op_8182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8182_end_0 = const()[name = string("op_8182_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8182_end_mask_0 = const()[name = string("op_8182_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8182_cast_fp16 = slice_by_index(begin = var_8182_begin_0, end = var_8182_end_0, end_mask = var_8182_end_mask_0, x = var_8172_cast_fp16)[name = string("op_8182_cast_fp16")];
            bool var_8191_transpose_x_1 = const()[name = string("op_8191_transpose_x_1"), val = bool(false)];
            bool var_8191_transpose_y_1 = const()[name = string("op_8191_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8191_cast_fp16 = matmul(transpose_x = var_8191_transpose_x_1, transpose_y = var_8191_transpose_y_1, x = var_8182_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<int32, [4]> var_8194_perm_0 = const()[name = string("op_8194_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8195_to_fp16 = const()[name = string("op_8195_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8194_cast_fp16 = transpose(perm = var_8194_perm_0, x = var_8191_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 2048]> var_8196_cast_fp16 = mul(x = var_8194_cast_fp16, y = var_8195_to_fp16)[name = string("op_8196_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_45_cast_fp16 = add(x = var_8196_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8202_transpose_x_1 = const()[name = string("op_8202_transpose_x_1"), val = bool(false)];
            bool var_8202_transpose_y_1 = const()[name = string("op_8202_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8202_cast_fp16 = matmul(transpose_x = var_8202_transpose_x_1, transpose_y = var_8202_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8202_cast_fp16")];
            fp16 var_8203_to_fp16 = const()[name = string("op_8203_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8204_cast_fp16 = mul(x = var_8202_cast_fp16, y = var_8203_to_fp16)[name = string("op_8204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8204_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8208 = const()[name = string("op_8208"), val = int32(-1)];
            bool var_8209_interleave_0 = const()[name = string("op_8209_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8209_cast_fp16 = concat(axis = var_8208, interleave = var_8209_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8209_cast_fp16")];
            int32 var_8210 = const()[name = string("op_8210"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_45_cast_fp16 = softmax(axis = var_8210, x = var_8209_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8217_begin_0 = const()[name = string("op_8217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8217_end_0 = const()[name = string("op_8217_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8217_end_mask_0 = const()[name = string("op_8217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8217_cast_fp16 = slice_by_index(begin = var_8217_begin_0, end = var_8217_end_0, end_mask = var_8217_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8217_cast_fp16")];
            tensor<int32, [4]> var_8222_begin_0 = const()[name = string("op_8222_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8222_end_0 = const()[name = string("op_8222_end_0"), val = tensor<int32, [4]>([23, 8, 2064, 128])];
            tensor<bool, [4]> var_8222_end_mask_0 = const()[name = string("op_8222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8222_cast_fp16 = slice_by_index(begin = var_8222_begin_0, end = var_8222_end_0, end_mask = var_8222_end_mask_0, x = read_state_1)[name = string("op_8222_cast_fp16")];
            tensor<int32, [4]> var_8232_begin_0 = const()[name = string("op_8232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8232_end_0 = const()[name = string("op_8232_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8232_end_mask_0 = const()[name = string("op_8232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8232_cast_fp16 = slice_by_index(begin = var_8232_begin_0, end = var_8232_end_0, end_mask = var_8232_end_mask_0, x = var_8222_cast_fp16)[name = string("op_8232_cast_fp16")];
            bool var_8238_transpose_x_0 = const()[name = string("op_8238_transpose_x_0"), val = bool(false)];
            bool var_8238_transpose_y_0 = const()[name = string("op_8238_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8238_cast_fp16 = matmul(transpose_x = var_8238_transpose_x_0, transpose_y = var_8238_transpose_y_0, x = var_8217_cast_fp16, y = var_8232_cast_fp16)[name = string("op_8238_cast_fp16")];
            tensor<int32, [4]> var_8243_begin_0 = const()[name = string("op_8243_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8243_end_0 = const()[name = string("op_8243_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8243_end_mask_0 = const()[name = string("op_8243_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8243_cast_fp16 = slice_by_index(begin = var_8243_begin_0, end = var_8243_end_0, end_mask = var_8243_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8243_cast_fp16")];
            bool var_8244_transpose_x_0 = const()[name = string("op_8244_transpose_x_0"), val = bool(false)];
            bool var_8244_transpose_y_0 = const()[name = string("op_8244_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8099_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8244_cast_fp16 = matmul(transpose_x = var_8244_transpose_x_0, transpose_y = var_8244_transpose_y_0, x = var_8243_cast_fp16, y = v_45_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8238_cast_fp16, y = var_8244_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8261_begin_0 = const()[name = string("op_8261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8261_end_0 = const()[name = string("op_8261_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8261_end_mask_0 = const()[name = string("op_8261_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8261_cast_fp16 = slice_by_index(begin = var_8261_begin_0, end = var_8261_end_0, end_mask = var_8261_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8261_cast_fp16")];
            tensor<int32, [4]> var_8276_begin_0 = const()[name = string("op_8276_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8276_end_0 = const()[name = string("op_8276_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8276_end_mask_0 = const()[name = string("op_8276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8276_cast_fp16 = slice_by_index(begin = var_8276_begin_0, end = var_8276_end_0, end_mask = var_8276_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8276_cast_fp16")];
            int32 var_8278 = const()[name = string("op_8278"), val = int32(1)];
            bool mixed_91_interleave_0 = const()[name = string("mixed_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_91_cast_fp16 = concat(axis = var_8278, interleave = mixed_91_interleave_0, values = (var_8261_cast_fp16, var_8276_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8282_perm_0 = const()[name = string("op_8282_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8286 = const()[name = string("op_8286"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8282_cast_fp16 = transpose(perm = var_8282_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8286, x = var_8282_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662151744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664248960))))[name = string("model_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_275_cast_fp16 = add(x = x_265_cast_fp16, y = linear_157_cast_fp16)[name = string("x_275_cast_fp16")];
            fp16 var_8297_promoted_to_fp16 = const()[name = string("op_8297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8299_cast_fp16 = pow(x = x_275_cast_fp16, y = var_8297_promoted_to_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [1]> var_8301_axes_0 = const()[name = string("op_8301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8301_keep_dims_0 = const()[name = string("op_8301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = reduce_mean(axes = var_8301_axes_0, keep_dims = var_8301_keep_dims_0, x = var_8299_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp16 var_8302_to_fp16 = const()[name = string("op_8302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8303_cast_fp16 = add(x = var_8301_cast_fp16, y = var_8302_to_fp16)[name = string("op_8303_cast_fp16")];
            fp32 var_8304_epsilon_0 = const()[name = string("op_8304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8304_cast_fp16 = rsqrt(epsilon = var_8304_epsilon_0, x = var_8303_cast_fp16)[name = string("op_8304_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8305_cast_fp16 = mul(x = x_275_cast_fp16, y = var_8304_cast_fp16)[name = string("op_8305_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664251072)))];
            tensor<fp16, [1, 16, 1024]> input_225_cast_fp16 = mul(x = var_8305_cast_fp16, y = model_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664253184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667398976))))[name = string("model_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8310_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8310_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667405184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670550976))))[name = string("model_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_229_cast_fp16 = mul(x = var_8310_cast_fp16, y = linear_159_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670557184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673702976))))[name = string("model_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_160_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_277_cast_fp16 = add(x = x_275_cast_fp16, y = linear_160_cast_fp16)[name = string("x_277_cast_fp16")];
            fp16 var_8324_promoted_to_fp16 = const()[name = string("op_8324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8326_cast_fp16 = pow(x = x_277_cast_fp16, y = var_8324_promoted_to_fp16)[name = string("op_8326_cast_fp16")];
            tensor<int32, [1]> var_8328_axes_0 = const()[name = string("op_8328_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8328_keep_dims_0 = const()[name = string("op_8328_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8328_cast_fp16 = reduce_mean(axes = var_8328_axes_0, keep_dims = var_8328_keep_dims_0, x = var_8326_cast_fp16)[name = string("op_8328_cast_fp16")];
            fp16 var_8329_to_fp16 = const()[name = string("op_8329_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = add(x = var_8328_cast_fp16, y = var_8329_to_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8332_cast_fp16 = mul(x = x_277_cast_fp16, y = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673705088)))];
            tensor<fp16, [1, 16, 1024]> input_231_cast_fp16 = mul(x = var_8332_cast_fp16, y = model_layers_23_input_layernorm_weight_to_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673707200))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675804416))))[name = string("model_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_8341 = const()[name = string("op_8341"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8342_cast_fp16 = reshape(shape = var_8341, x = linear_161_cast_fp16)[name = string("op_8342_cast_fp16")];
            tensor<int32, [4]> x_279_perm_0 = const()[name = string("x_279_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8350_promoted_to_fp16 = const()[name = string("op_8350_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_279_cast_fp16 = transpose(perm = x_279_perm_0, x = var_8342_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_8352_cast_fp16 = pow(x = x_279_cast_fp16, y = var_8350_promoted_to_fp16)[name = string("op_8352_cast_fp16")];
            tensor<int32, [1]> var_8354_axes_0 = const()[name = string("op_8354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8354_keep_dims_0 = const()[name = string("op_8354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8354_cast_fp16 = reduce_mean(axes = var_8354_axes_0, keep_dims = var_8354_keep_dims_0, x = var_8352_cast_fp16)[name = string("op_8354_cast_fp16")];
            fp16 var_8355_to_fp16 = const()[name = string("op_8355_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8356_cast_fp16 = add(x = var_8354_cast_fp16, y = var_8355_to_fp16)[name = string("op_8356_cast_fp16")];
            fp32 var_8357_epsilon_0 = const()[name = string("op_8357_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8357_cast_fp16 = rsqrt(epsilon = var_8357_epsilon_0, x = var_8356_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8358_cast_fp16 = mul(x = x_279_cast_fp16, y = var_8357_cast_fp16)[name = string("op_8358_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808576)))];
            tensor<fp16, [1, 16, 16, 128]> x_283_cast_fp16 = mul(x = var_8358_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676857536))))[name = string("model_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_8367 = const()[name = string("op_8367"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8368_cast_fp16 = reshape(shape = var_8367, x = linear_162_cast_fp16)[name = string("op_8368_cast_fp16")];
            tensor<int32, [4]> x_281_perm_0 = const()[name = string("x_281_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8376_promoted_to_fp16 = const()[name = string("op_8376_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_281_cast_fp16 = transpose(perm = x_281_perm_0, x = var_8368_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_8378_cast_fp16 = pow(x = x_281_cast_fp16, y = var_8376_promoted_to_fp16)[name = string("op_8378_cast_fp16")];
            tensor<int32, [1]> var_8380_axes_0 = const()[name = string("op_8380_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8380_keep_dims_0 = const()[name = string("op_8380_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8380_cast_fp16 = reduce_mean(axes = var_8380_axes_0, keep_dims = var_8380_keep_dims_0, x = var_8378_cast_fp16)[name = string("op_8380_cast_fp16")];
            fp16 var_8381_to_fp16 = const()[name = string("op_8381_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8382_cast_fp16 = add(x = var_8380_cast_fp16, y = var_8381_to_fp16)[name = string("op_8382_cast_fp16")];
            fp32 var_8383_epsilon_0 = const()[name = string("op_8383_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8383_cast_fp16 = rsqrt(epsilon = var_8383_epsilon_0, x = var_8382_cast_fp16)[name = string("op_8383_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8384_cast_fp16 = mul(x = x_281_cast_fp16, y = var_8383_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859648)))];
            tensor<fp16, [1, 8, 16, 128]> x_285_cast_fp16 = mul(x = var_8384_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677908608))))[name = string("model_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8393 = const()[name = string("op_8393"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8394_cast_fp16 = reshape(shape = var_8393, x = linear_163_cast_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_93_begin_0 = const()[name = string("first_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_93_end_0 = const()[name = string("first_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_93_end_mask_0 = const()[name = string("first_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_93_cast_fp16 = slice_by_index(begin = first_93_begin_0, end = first_93_end_0, end_mask = first_93_end_mask_0, x = x_283_cast_fp16)[name = string("first_93_cast_fp16")];
            tensor<int32, [4]> second_93_begin_0 = const()[name = string("second_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_93_end_0 = const()[name = string("second_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_93_end_mask_0 = const()[name = string("second_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_93_cast_fp16 = slice_by_index(begin = second_93_begin_0, end = second_93_end_0, end_mask = second_93_end_mask_0, x = x_283_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8408_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8408_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8409_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8409_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8411_cast_fp16 = sub(x = var_8408_cast_fp16, y = var_8409_cast_fp16)[name = string("op_8411_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8412_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8412_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8413_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8415_cast_fp16 = add(x = var_8412_cast_fp16, y = var_8413_cast_fp16)[name = string("op_8415_cast_fp16")];
            int32 var_8417 = const()[name = string("op_8417"), val = int32(-1)];
            bool q_47_interleave_0 = const()[name = string("q_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_47_cast_fp16 = concat(axis = var_8417, interleave = q_47_interleave_0, values = (var_8411_cast_fp16, var_8415_cast_fp16))[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> first_95_begin_0 = const()[name = string("first_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_95_end_0 = const()[name = string("first_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_95_end_mask_0 = const()[name = string("first_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_95_cast_fp16 = slice_by_index(begin = first_95_begin_0, end = first_95_end_0, end_mask = first_95_end_mask_0, x = x_285_cast_fp16)[name = string("first_95_cast_fp16")];
            tensor<int32, [4]> second_95_begin_0 = const()[name = string("second_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_95_end_0 = const()[name = string("second_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_95_end_mask_0 = const()[name = string("second_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_95_cast_fp16 = slice_by_index(begin = second_95_begin_0, end = second_95_end_0, end_mask = second_95_end_mask_0, x = x_285_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8429_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8429_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8430_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8432_cast_fp16 = sub(x = var_8429_cast_fp16, y = var_8430_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8433_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8434_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8434_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8436_cast_fp16 = add(x = var_8433_cast_fp16, y = var_8434_cast_fp16)[name = string("op_8436_cast_fp16")];
            int32 var_8438 = const()[name = string("op_8438"), val = int32(-1)];
            bool k_47_interleave_0 = const()[name = string("k_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_47_cast_fp16 = concat(axis = var_8438, interleave = k_47_interleave_0, values = (var_8432_cast_fp16, var_8436_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_8449_begin_0 = const()[name = string("op_8449_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8449_end_0 = const()[name = string("op_8449_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8449_end_mask_0 = const()[name = string("op_8449_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8449_stride_0 = const()[name = string("op_8449_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8449_cast_fp16 = slice_by_index(begin = var_8449_begin_0, end = var_8449_end_0, end_mask = var_8449_end_mask_0, stride = var_8449_stride_0, x = q_47_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<int32, [4]> var_8459_begin_0 = const()[name = string("op_8459_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8459_end_0 = const()[name = string("op_8459_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8459_end_mask_0 = const()[name = string("op_8459_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8459_stride_0 = const()[name = string("op_8459_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8459_cast_fp16 = slice_by_index(begin = var_8459_begin_0, end = var_8459_end_0, end_mask = var_8459_end_mask_0, stride = var_8459_stride_0, x = q_47_cast_fp16)[name = string("op_8459_cast_fp16")];
            int32 var_8461 = const()[name = string("op_8461"), val = int32(2)];
            bool stacked_47_interleave_0 = const()[name = string("stacked_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_47_cast_fp16 = concat(axis = var_8461, interleave = stacked_47_interleave_0, values = (var_8449_cast_fp16, var_8459_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_8467_begin_0 = const()[name = string("op_8467_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8467_end_0 = const()[name = string("op_8467_end_0"), val = tensor<int32, [4]>([24, 8, 2064, 128])];
            tensor<bool, [4]> var_8467_end_mask_0 = const()[name = string("op_8467_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8467_cast_fp16 = slice_by_index(begin = var_8467_begin_0, end = var_8467_end_0, end_mask = var_8467_end_mask_0, x = read_state_0)[name = string("op_8467_cast_fp16")];
            tensor<int32, [4]> var_8477_begin_0 = const()[name = string("op_8477_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8477_end_0 = const()[name = string("op_8477_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8477_end_mask_0 = const()[name = string("op_8477_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8477_cast_fp16 = slice_by_index(begin = var_8477_begin_0, end = var_8477_end_0, end_mask = var_8477_end_mask_0, x = var_8467_cast_fp16)[name = string("op_8477_cast_fp16")];
            bool var_8486_transpose_x_1 = const()[name = string("op_8486_transpose_x_1"), val = bool(false)];
            bool var_8486_transpose_y_1 = const()[name = string("op_8486_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8486_cast_fp16 = matmul(transpose_x = var_8486_transpose_x_1, transpose_y = var_8486_transpose_y_1, x = var_8477_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8489_perm_0 = const()[name = string("op_8489_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8490_to_fp16 = const()[name = string("op_8490_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8489_cast_fp16 = transpose(perm = var_8489_perm_0, x = var_8486_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 2048]> var_8491_cast_fp16 = mul(x = var_8489_cast_fp16, y = var_8490_to_fp16)[name = string("op_8491_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_47_cast_fp16 = add(x = var_8491_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_8497_transpose_x_1 = const()[name = string("op_8497_transpose_x_1"), val = bool(false)];
            bool var_8497_transpose_y_1 = const()[name = string("op_8497_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8497_cast_fp16 = matmul(transpose_x = var_8497_transpose_x_1, transpose_y = var_8497_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_8497_cast_fp16")];
            fp16 var_8498_to_fp16 = const()[name = string("op_8498_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8499_cast_fp16 = mul(x = var_8497_cast_fp16, y = var_8498_to_fp16)[name = string("op_8499_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_8499_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_8503 = const()[name = string("op_8503"), val = int32(-1)];
            bool var_8504_interleave_0 = const()[name = string("op_8504_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8504_cast_fp16 = concat(axis = var_8503, interleave = var_8504_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_8504_cast_fp16")];
            int32 var_8505 = const()[name = string("op_8505"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_47_cast_fp16 = softmax(axis = var_8505, x = var_8504_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_8512_begin_0 = const()[name = string("op_8512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8512_end_0 = const()[name = string("op_8512_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8512_end_mask_0 = const()[name = string("op_8512_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8512_cast_fp16 = slice_by_index(begin = var_8512_begin_0, end = var_8512_end_0, end_mask = var_8512_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8512_cast_fp16")];
            tensor<int32, [4]> var_8517_begin_0 = const()[name = string("op_8517_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8517_end_0 = const()[name = string("op_8517_end_0"), val = tensor<int32, [4]>([24, 8, 2064, 128])];
            tensor<bool, [4]> var_8517_end_mask_0 = const()[name = string("op_8517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8517_cast_fp16 = slice_by_index(begin = var_8517_begin_0, end = var_8517_end_0, end_mask = var_8517_end_mask_0, x = read_state_1)[name = string("op_8517_cast_fp16")];
            tensor<int32, [4]> var_8527_begin_0 = const()[name = string("op_8527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8527_end_0 = const()[name = string("op_8527_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8527_end_mask_0 = const()[name = string("op_8527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8527_cast_fp16 = slice_by_index(begin = var_8527_begin_0, end = var_8527_end_0, end_mask = var_8527_end_mask_0, x = var_8517_cast_fp16)[name = string("op_8527_cast_fp16")];
            bool var_8533_transpose_x_0 = const()[name = string("op_8533_transpose_x_0"), val = bool(false)];
            bool var_8533_transpose_y_0 = const()[name = string("op_8533_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8533_cast_fp16 = matmul(transpose_x = var_8533_transpose_x_0, transpose_y = var_8533_transpose_y_0, x = var_8512_cast_fp16, y = var_8527_cast_fp16)[name = string("op_8533_cast_fp16")];
            tensor<int32, [4]> var_8538_begin_0 = const()[name = string("op_8538_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8538_end_0 = const()[name = string("op_8538_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8538_end_mask_0 = const()[name = string("op_8538_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8538_cast_fp16 = slice_by_index(begin = var_8538_begin_0, end = var_8538_end_0, end_mask = var_8538_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8538_cast_fp16")];
            bool var_8539_transpose_x_0 = const()[name = string("op_8539_transpose_x_0"), val = bool(false)];
            bool var_8539_transpose_y_0 = const()[name = string("op_8539_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_8394_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_8539_cast_fp16 = matmul(transpose_x = var_8539_transpose_x_0, transpose_y = var_8539_transpose_y_0, x = var_8538_cast_fp16, y = v_47_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_8533_cast_fp16, y = var_8539_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_8556_begin_0 = const()[name = string("op_8556_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8556_end_0 = const()[name = string("op_8556_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8556_end_mask_0 = const()[name = string("op_8556_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8556_cast_fp16 = slice_by_index(begin = var_8556_begin_0, end = var_8556_end_0, end_mask = var_8556_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8556_cast_fp16")];
            tensor<int32, [4]> var_8571_begin_0 = const()[name = string("op_8571_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8571_end_0 = const()[name = string("op_8571_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8571_end_mask_0 = const()[name = string("op_8571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8571_cast_fp16 = slice_by_index(begin = var_8571_begin_0, end = var_8571_end_0, end_mask = var_8571_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8571_cast_fp16")];
            int32 var_8573 = const()[name = string("op_8573"), val = int32(1)];
            bool mixed_95_interleave_0 = const()[name = string("mixed_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_95_cast_fp16 = concat(axis = var_8573, interleave = mixed_95_interleave_0, values = (var_8556_cast_fp16, var_8571_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_8577_perm_0 = const()[name = string("op_8577_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8581 = const()[name = string("op_8581"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8577_cast_fp16 = transpose(perm = var_8577_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_8581, x = var_8577_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677910720))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680007936))))[name = string("model_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_164_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_287_cast_fp16 = add(x = x_277_cast_fp16, y = linear_164_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_8592_promoted_to_fp16 = const()[name = string("op_8592_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8594_cast_fp16 = pow(x = x_287_cast_fp16, y = var_8592_promoted_to_fp16)[name = string("op_8594_cast_fp16")];
            tensor<int32, [1]> var_8596_axes_0 = const()[name = string("op_8596_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8596_keep_dims_0 = const()[name = string("op_8596_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8596_cast_fp16 = reduce_mean(axes = var_8596_axes_0, keep_dims = var_8596_keep_dims_0, x = var_8594_cast_fp16)[name = string("op_8596_cast_fp16")];
            fp16 var_8597_to_fp16 = const()[name = string("op_8597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8598_cast_fp16 = add(x = var_8596_cast_fp16, y = var_8597_to_fp16)[name = string("op_8598_cast_fp16")];
            fp32 var_8599_epsilon_0 = const()[name = string("op_8599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8599_cast_fp16 = rsqrt(epsilon = var_8599_epsilon_0, x = var_8598_cast_fp16)[name = string("op_8599_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8600_cast_fp16 = mul(x = x_287_cast_fp16, y = var_8599_cast_fp16)[name = string("op_8600_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680010048)))];
            tensor<fp16, [1, 16, 1024]> input_235_cast_fp16 = mul(x = var_8600_cast_fp16, y = model_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680012160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683157952))))[name = string("model_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8605_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_8605_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683164160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686309952))))[name = string("model_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_239_cast_fp16 = mul(x = var_8605_cast_fp16, y = linear_166_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686316160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689461952))))[name = string("model_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_167_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_289_cast_fp16 = add(x = x_287_cast_fp16, y = linear_167_cast_fp16)[name = string("x_289_cast_fp16")];
            fp16 var_8619_promoted_to_fp16 = const()[name = string("op_8619_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8621_cast_fp16 = pow(x = x_289_cast_fp16, y = var_8619_promoted_to_fp16)[name = string("op_8621_cast_fp16")];
            tensor<int32, [1]> var_8623_axes_0 = const()[name = string("op_8623_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8623_keep_dims_0 = const()[name = string("op_8623_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = reduce_mean(axes = var_8623_axes_0, keep_dims = var_8623_keep_dims_0, x = var_8621_cast_fp16)[name = string("op_8623_cast_fp16")];
            fp16 var_8624_to_fp16 = const()[name = string("op_8624_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_8624_to_fp16)[name = string("op_8625_cast_fp16")];
            fp32 var_8626_epsilon_0 = const()[name = string("op_8626_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8626_cast_fp16 = rsqrt(epsilon = var_8626_epsilon_0, x = var_8625_cast_fp16)[name = string("op_8626_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8627_cast_fp16 = mul(x = x_289_cast_fp16, y = var_8626_cast_fp16)[name = string("op_8627_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689464064)))];
            tensor<fp16, [1, 16, 1024]> input_241_cast_fp16 = mul(x = var_8627_cast_fp16, y = model_layers_24_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_24_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689466176))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691563392))))[name = string("model_layers_24_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_24_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [4]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8637_cast_fp16 = reshape(shape = var_8636, x = linear_168_cast_fp16)[name = string("op_8637_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8645_promoted_to_fp16 = const()[name = string("op_8645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_8637_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_8647_cast_fp16 = pow(x = x_291_cast_fp16, y = var_8645_promoted_to_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> var_8649_axes_0 = const()[name = string("op_8649_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8649_keep_dims_0 = const()[name = string("op_8649_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8649_cast_fp16 = reduce_mean(axes = var_8649_axes_0, keep_dims = var_8649_keep_dims_0, x = var_8647_cast_fp16)[name = string("op_8649_cast_fp16")];
            fp16 var_8650_to_fp16 = const()[name = string("op_8650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8651_cast_fp16 = add(x = var_8649_cast_fp16, y = var_8650_to_fp16)[name = string("op_8651_cast_fp16")];
            fp32 var_8652_epsilon_0 = const()[name = string("op_8652_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8652_cast_fp16 = rsqrt(epsilon = var_8652_epsilon_0, x = var_8651_cast_fp16)[name = string("op_8652_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8653_cast_fp16 = mul(x = x_291_cast_fp16, y = var_8652_cast_fp16)[name = string("op_8653_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567552)))];
            tensor<fp16, [1, 16, 16, 128]> x_295_cast_fp16 = mul(x = var_8653_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_295_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692616512))))[name = string("model_layers_24_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [4]> var_8662 = const()[name = string("op_8662"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8663_cast_fp16 = reshape(shape = var_8662, x = linear_169_cast_fp16)[name = string("op_8663_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8671_promoted_to_fp16 = const()[name = string("op_8671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_8663_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_8673_cast_fp16 = pow(x = x_293_cast_fp16, y = var_8671_promoted_to_fp16)[name = string("op_8673_cast_fp16")];
            tensor<int32, [1]> var_8675_axes_0 = const()[name = string("op_8675_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8675_keep_dims_0 = const()[name = string("op_8675_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8675_cast_fp16 = reduce_mean(axes = var_8675_axes_0, keep_dims = var_8675_keep_dims_0, x = var_8673_cast_fp16)[name = string("op_8675_cast_fp16")];
            fp16 var_8676_to_fp16 = const()[name = string("op_8676_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8677_cast_fp16 = add(x = var_8675_cast_fp16, y = var_8676_to_fp16)[name = string("op_8677_cast_fp16")];
            fp32 var_8678_epsilon_0 = const()[name = string("op_8678_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8678_cast_fp16 = rsqrt(epsilon = var_8678_epsilon_0, x = var_8677_cast_fp16)[name = string("op_8678_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8679_cast_fp16 = mul(x = x_293_cast_fp16, y = var_8678_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618624)))];
            tensor<fp16, [1, 8, 16, 128]> x_297_cast_fp16 = mul(x = var_8679_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_297_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693667584))))[name = string("model_layers_24_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> var_8688 = const()[name = string("op_8688"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8689_cast_fp16 = reshape(shape = var_8688, x = linear_170_cast_fp16)[name = string("op_8689_cast_fp16")];
            tensor<int32, [4]> v_49_perm_0 = const()[name = string("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_97_begin_0 = const()[name = string("first_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_97_end_0 = const()[name = string("first_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_97_end_mask_0 = const()[name = string("first_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_97_cast_fp16 = slice_by_index(begin = first_97_begin_0, end = first_97_end_0, end_mask = first_97_end_mask_0, x = x_295_cast_fp16)[name = string("first_97_cast_fp16")];
            tensor<int32, [4]> second_97_begin_0 = const()[name = string("second_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_97_end_0 = const()[name = string("second_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_97_end_mask_0 = const()[name = string("second_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_97_cast_fp16 = slice_by_index(begin = second_97_begin_0, end = second_97_end_0, end_mask = second_97_end_mask_0, x = x_295_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8703_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8704_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8704_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8706_cast_fp16 = sub(x = var_8703_cast_fp16, y = var_8704_cast_fp16)[name = string("op_8706_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8707_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8707_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8708_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8708_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8710_cast_fp16 = add(x = var_8707_cast_fp16, y = var_8708_cast_fp16)[name = string("op_8710_cast_fp16")];
            int32 var_8712 = const()[name = string("op_8712"), val = int32(-1)];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_49_cast_fp16 = concat(axis = var_8712, interleave = q_49_interleave_0, values = (var_8706_cast_fp16, var_8710_cast_fp16))[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> first_99_begin_0 = const()[name = string("first_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_99_end_0 = const()[name = string("first_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_99_end_mask_0 = const()[name = string("first_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_99_cast_fp16 = slice_by_index(begin = first_99_begin_0, end = first_99_end_0, end_mask = first_99_end_mask_0, x = x_297_cast_fp16)[name = string("first_99_cast_fp16")];
            tensor<int32, [4]> second_99_begin_0 = const()[name = string("second_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_99_end_0 = const()[name = string("second_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_99_end_mask_0 = const()[name = string("second_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_99_cast_fp16 = slice_by_index(begin = second_99_begin_0, end = second_99_end_0, end_mask = second_99_end_mask_0, x = x_297_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8724_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8724_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8725_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8727_cast_fp16 = sub(x = var_8724_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8727_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8728_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8729_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8729_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8731_cast_fp16 = add(x = var_8728_cast_fp16, y = var_8729_cast_fp16)[name = string("op_8731_cast_fp16")];
            int32 var_8733 = const()[name = string("op_8733"), val = int32(-1)];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_49_cast_fp16 = concat(axis = var_8733, interleave = k_49_interleave_0, values = (var_8727_cast_fp16, var_8731_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_8744_begin_0 = const()[name = string("op_8744_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8744_end_0 = const()[name = string("op_8744_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8744_end_mask_0 = const()[name = string("op_8744_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8744_stride_0 = const()[name = string("op_8744_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, stride = var_8744_stride_0, x = q_49_cast_fp16)[name = string("op_8744_cast_fp16")];
            tensor<int32, [4]> var_8754_begin_0 = const()[name = string("op_8754_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8754_end_0 = const()[name = string("op_8754_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8754_end_mask_0 = const()[name = string("op_8754_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8754_stride_0 = const()[name = string("op_8754_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8754_cast_fp16 = slice_by_index(begin = var_8754_begin_0, end = var_8754_end_0, end_mask = var_8754_end_mask_0, stride = var_8754_stride_0, x = q_49_cast_fp16)[name = string("op_8754_cast_fp16")];
            int32 var_8756 = const()[name = string("op_8756"), val = int32(2)];
            bool stacked_49_interleave_0 = const()[name = string("stacked_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_49_cast_fp16 = concat(axis = var_8756, interleave = stacked_49_interleave_0, values = (var_8744_cast_fp16, var_8754_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_8762_begin_0 = const()[name = string("op_8762_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8762_end_0 = const()[name = string("op_8762_end_0"), val = tensor<int32, [4]>([25, 8, 2064, 128])];
            tensor<bool, [4]> var_8762_end_mask_0 = const()[name = string("op_8762_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8762_cast_fp16 = slice_by_index(begin = var_8762_begin_0, end = var_8762_end_0, end_mask = var_8762_end_mask_0, x = read_state_0)[name = string("op_8762_cast_fp16")];
            tensor<int32, [4]> var_8772_begin_0 = const()[name = string("op_8772_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8772_end_0 = const()[name = string("op_8772_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8772_end_mask_0 = const()[name = string("op_8772_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8772_cast_fp16 = slice_by_index(begin = var_8772_begin_0, end = var_8772_end_0, end_mask = var_8772_end_mask_0, x = var_8762_cast_fp16)[name = string("op_8772_cast_fp16")];
            bool var_8781_transpose_x_1 = const()[name = string("op_8781_transpose_x_1"), val = bool(false)];
            bool var_8781_transpose_y_1 = const()[name = string("op_8781_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_8781_cast_fp16 = matmul(transpose_x = var_8781_transpose_x_1, transpose_y = var_8781_transpose_y_1, x = var_8772_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_8781_cast_fp16")];
            tensor<int32, [4]> var_8784_perm_0 = const()[name = string("op_8784_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8785_to_fp16 = const()[name = string("op_8785_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_8784_cast_fp16 = transpose(perm = var_8784_perm_0, x = var_8781_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 2048]> var_8786_cast_fp16 = mul(x = var_8784_cast_fp16, y = var_8785_to_fp16)[name = string("op_8786_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_49_cast_fp16 = add(x = var_8786_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_8792_transpose_x_1 = const()[name = string("op_8792_transpose_x_1"), val = bool(false)];
            bool var_8792_transpose_y_1 = const()[name = string("op_8792_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8792_cast_fp16 = matmul(transpose_x = var_8792_transpose_x_1, transpose_y = var_8792_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_8792_cast_fp16")];
            fp16 var_8793_to_fp16 = const()[name = string("op_8793_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8794_cast_fp16 = mul(x = var_8792_cast_fp16, y = var_8793_to_fp16)[name = string("op_8794_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_8794_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_8798 = const()[name = string("op_8798"), val = int32(-1)];
            bool var_8799_interleave_0 = const()[name = string("op_8799_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_8799_cast_fp16 = concat(axis = var_8798, interleave = var_8799_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_8799_cast_fp16")];
            int32 var_8800 = const()[name = string("op_8800"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_49_cast_fp16 = softmax(axis = var_8800, x = var_8799_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_8807_begin_0 = const()[name = string("op_8807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8807_end_0 = const()[name = string("op_8807_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_8807_end_mask_0 = const()[name = string("op_8807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_8807_cast_fp16 = slice_by_index(begin = var_8807_begin_0, end = var_8807_end_0, end_mask = var_8807_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8807_cast_fp16")];
            tensor<int32, [4]> var_8812_begin_0 = const()[name = string("op_8812_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8812_end_0 = const()[name = string("op_8812_end_0"), val = tensor<int32, [4]>([25, 8, 2064, 128])];
            tensor<bool, [4]> var_8812_end_mask_0 = const()[name = string("op_8812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_8812_cast_fp16 = slice_by_index(begin = var_8812_begin_0, end = var_8812_end_0, end_mask = var_8812_end_mask_0, x = read_state_1)[name = string("op_8812_cast_fp16")];
            tensor<int32, [4]> var_8822_begin_0 = const()[name = string("op_8822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8822_end_0 = const()[name = string("op_8822_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_8822_end_mask_0 = const()[name = string("op_8822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_8822_cast_fp16 = slice_by_index(begin = var_8822_begin_0, end = var_8822_end_0, end_mask = var_8822_end_mask_0, x = var_8812_cast_fp16)[name = string("op_8822_cast_fp16")];
            bool var_8828_transpose_x_0 = const()[name = string("op_8828_transpose_x_0"), val = bool(false)];
            bool var_8828_transpose_y_0 = const()[name = string("op_8828_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_0, transpose_y = var_8828_transpose_y_0, x = var_8807_cast_fp16, y = var_8822_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8833_begin_0 = const()[name = string("op_8833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_8833_end_0 = const()[name = string("op_8833_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_8833_end_mask_0 = const()[name = string("op_8833_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8833_cast_fp16 = slice_by_index(begin = var_8833_begin_0, end = var_8833_end_0, end_mask = var_8833_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8833_cast_fp16")];
            bool var_8834_transpose_x_0 = const()[name = string("op_8834_transpose_x_0"), val = bool(false)];
            bool var_8834_transpose_y_0 = const()[name = string("op_8834_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_8689_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_8834_cast_fp16 = matmul(transpose_x = var_8834_transpose_x_0, transpose_y = var_8834_transpose_y_0, x = var_8833_cast_fp16, y = v_49_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_8828_cast_fp16, y = var_8834_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_8851_begin_0 = const()[name = string("op_8851_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8851_end_0 = const()[name = string("op_8851_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8851_end_mask_0 = const()[name = string("op_8851_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8851_cast_fp16 = slice_by_index(begin = var_8851_begin_0, end = var_8851_end_0, end_mask = var_8851_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<int32, [4]> var_8866_begin_0 = const()[name = string("op_8866_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8866_end_0 = const()[name = string("op_8866_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8866_end_mask_0 = const()[name = string("op_8866_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8866_cast_fp16 = slice_by_index(begin = var_8866_begin_0, end = var_8866_end_0, end_mask = var_8866_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8866_cast_fp16")];
            int32 var_8868 = const()[name = string("op_8868"), val = int32(1)];
            bool mixed_99_interleave_0 = const()[name = string("mixed_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_99_cast_fp16 = concat(axis = var_8868, interleave = mixed_99_interleave_0, values = (var_8851_cast_fp16, var_8866_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_8872_perm_0 = const()[name = string("op_8872_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8876 = const()[name = string("op_8876"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8872_cast_fp16 = transpose(perm = var_8872_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_8876, x = var_8872_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_24_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693669696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695766912))))[name = string("model_layers_24_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_171_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = linear_171_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_8887_promoted_to_fp16 = const()[name = string("op_8887_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8889_cast_fp16 = pow(x = x_299_cast_fp16, y = var_8887_promoted_to_fp16)[name = string("op_8889_cast_fp16")];
            tensor<int32, [1]> var_8891_axes_0 = const()[name = string("op_8891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8891_keep_dims_0 = const()[name = string("op_8891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8891_cast_fp16 = reduce_mean(axes = var_8891_axes_0, keep_dims = var_8891_keep_dims_0, x = var_8889_cast_fp16)[name = string("op_8891_cast_fp16")];
            fp16 var_8892_to_fp16 = const()[name = string("op_8892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8893_cast_fp16 = add(x = var_8891_cast_fp16, y = var_8892_to_fp16)[name = string("op_8893_cast_fp16")];
            fp32 var_8894_epsilon_0 = const()[name = string("op_8894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8894_cast_fp16 = rsqrt(epsilon = var_8894_epsilon_0, x = var_8893_cast_fp16)[name = string("op_8894_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8895_cast_fp16 = mul(x = x_299_cast_fp16, y = var_8894_cast_fp16)[name = string("op_8895_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695769024)))];
            tensor<fp16, [1, 16, 1024]> input_245_cast_fp16 = mul(x = var_8895_cast_fp16, y = model_layers_24_post_attention_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695771136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698916928))))[name = string("model_layers_24_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8900_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_8900_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698923136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702068928))))[name = string("model_layers_24_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_249_cast_fp16 = mul(x = var_8900_cast_fp16, y = linear_173_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_24_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702075136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705220928))))[name = string("model_layers_24_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_174_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_mlp_down_proj_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_301_cast_fp16 = add(x = x_299_cast_fp16, y = linear_174_cast_fp16)[name = string("x_301_cast_fp16")];
            fp16 var_8914_promoted_to_fp16 = const()[name = string("op_8914_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8916_cast_fp16 = pow(x = x_301_cast_fp16, y = var_8914_promoted_to_fp16)[name = string("op_8916_cast_fp16")];
            tensor<int32, [1]> var_8918_axes_0 = const()[name = string("op_8918_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8918_keep_dims_0 = const()[name = string("op_8918_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8918_cast_fp16 = reduce_mean(axes = var_8918_axes_0, keep_dims = var_8918_keep_dims_0, x = var_8916_cast_fp16)[name = string("op_8918_cast_fp16")];
            fp16 var_8919_to_fp16 = const()[name = string("op_8919_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8920_cast_fp16 = add(x = var_8918_cast_fp16, y = var_8919_to_fp16)[name = string("op_8920_cast_fp16")];
            fp32 var_8921_epsilon_0 = const()[name = string("op_8921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8921_cast_fp16 = rsqrt(epsilon = var_8921_epsilon_0, x = var_8920_cast_fp16)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8922_cast_fp16 = mul(x = x_301_cast_fp16, y = var_8921_cast_fp16)[name = string("op_8922_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705223040)))];
            tensor<fp16, [1, 16, 1024]> input_251_cast_fp16 = mul(x = var_8922_cast_fp16, y = model_layers_25_input_layernorm_weight_to_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_25_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705225152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707322368))))[name = string("model_layers_25_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_25_self_attn_q_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> var_8931 = const()[name = string("op_8931"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8932_cast_fp16 = reshape(shape = var_8931, x = linear_175_cast_fp16)[name = string("op_8932_cast_fp16")];
            tensor<int32, [4]> x_303_perm_0 = const()[name = string("x_303_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8940_promoted_to_fp16 = const()[name = string("op_8940_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_303_cast_fp16 = transpose(perm = x_303_perm_0, x = var_8932_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_8942_cast_fp16 = pow(x = x_303_cast_fp16, y = var_8940_promoted_to_fp16)[name = string("op_8942_cast_fp16")];
            tensor<int32, [1]> var_8944_axes_0 = const()[name = string("op_8944_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8944_keep_dims_0 = const()[name = string("op_8944_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8944_cast_fp16 = reduce_mean(axes = var_8944_axes_0, keep_dims = var_8944_keep_dims_0, x = var_8942_cast_fp16)[name = string("op_8944_cast_fp16")];
            fp16 var_8945_to_fp16 = const()[name = string("op_8945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8946_cast_fp16 = add(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            fp32 var_8947_epsilon_0 = const()[name = string("op_8947_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8947_cast_fp16 = rsqrt(epsilon = var_8947_epsilon_0, x = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8948_cast_fp16 = mul(x = x_303_cast_fp16, y = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326528)))];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = mul(x = var_8948_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708375488))))[name = string("model_layers_25_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_176_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_k_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> var_8957 = const()[name = string("op_8957"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8958_cast_fp16 = reshape(shape = var_8957, x = linear_176_cast_fp16)[name = string("op_8958_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_8958_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_8968_cast_fp16 = pow(x = x_305_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8968_cast_fp16")];
            tensor<int32, [1]> var_8970_axes_0 = const()[name = string("op_8970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8970_keep_dims_0 = const()[name = string("op_8970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8970_cast_fp16 = reduce_mean(axes = var_8970_axes_0, keep_dims = var_8970_keep_dims_0, x = var_8968_cast_fp16)[name = string("op_8970_cast_fp16")];
            fp16 var_8971_to_fp16 = const()[name = string("op_8971_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8972_cast_fp16 = add(x = var_8970_cast_fp16, y = var_8971_to_fp16)[name = string("op_8972_cast_fp16")];
            fp32 var_8973_epsilon_0 = const()[name = string("op_8973_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8973_cast_fp16 = rsqrt(epsilon = var_8973_epsilon_0, x = var_8972_cast_fp16)[name = string("op_8973_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8974_cast_fp16 = mul(x = x_305_cast_fp16, y = var_8973_cast_fp16)[name = string("op_8974_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377600)))];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = mul(x = var_8974_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_309_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709426560))))[name = string("model_layers_25_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_v_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> var_8983 = const()[name = string("op_8983"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8984_cast_fp16 = reshape(shape = var_8983, x = linear_177_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_101_begin_0 = const()[name = string("first_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_101_end_0 = const()[name = string("first_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_101_end_mask_0 = const()[name = string("first_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_101_cast_fp16 = slice_by_index(begin = first_101_begin_0, end = first_101_end_0, end_mask = first_101_end_mask_0, x = x_307_cast_fp16)[name = string("first_101_cast_fp16")];
            tensor<int32, [4]> second_101_begin_0 = const()[name = string("second_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_101_end_0 = const()[name = string("second_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_101_end_mask_0 = const()[name = string("second_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_101_cast_fp16 = slice_by_index(begin = second_101_begin_0, end = second_101_end_0, end_mask = second_101_end_mask_0, x = x_307_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8998_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8998_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8999_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8999_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9001_cast_fp16 = sub(x = var_8998_cast_fp16, y = var_8999_cast_fp16)[name = string("op_9001_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9002_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9002_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9003_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9005_cast_fp16 = add(x = var_9002_cast_fp16, y = var_9003_cast_fp16)[name = string("op_9005_cast_fp16")];
            int32 var_9007 = const()[name = string("op_9007"), val = int32(-1)];
            bool q_51_interleave_0 = const()[name = string("q_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_51_cast_fp16 = concat(axis = var_9007, interleave = q_51_interleave_0, values = (var_9001_cast_fp16, var_9005_cast_fp16))[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> first_103_begin_0 = const()[name = string("first_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_103_end_0 = const()[name = string("first_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_103_end_mask_0 = const()[name = string("first_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_103_cast_fp16 = slice_by_index(begin = first_103_begin_0, end = first_103_end_0, end_mask = first_103_end_mask_0, x = x_309_cast_fp16)[name = string("first_103_cast_fp16")];
            tensor<int32, [4]> second_103_begin_0 = const()[name = string("second_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_103_end_0 = const()[name = string("second_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_103_end_mask_0 = const()[name = string("second_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_103_cast_fp16 = slice_by_index(begin = second_103_begin_0, end = second_103_end_0, end_mask = second_103_end_mask_0, x = x_309_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9019_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9019_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9020_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9020_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9022_cast_fp16 = sub(x = var_9019_cast_fp16, y = var_9020_cast_fp16)[name = string("op_9022_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9023_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9024_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9024_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9026_cast_fp16 = add(x = var_9023_cast_fp16, y = var_9024_cast_fp16)[name = string("op_9026_cast_fp16")];
            int32 var_9028 = const()[name = string("op_9028"), val = int32(-1)];
            bool k_51_interleave_0 = const()[name = string("k_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_51_cast_fp16 = concat(axis = var_9028, interleave = k_51_interleave_0, values = (var_9022_cast_fp16, var_9026_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9039_begin_0 = const()[name = string("op_9039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9039_end_0 = const()[name = string("op_9039_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9039_end_mask_0 = const()[name = string("op_9039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9039_stride_0 = const()[name = string("op_9039_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9039_cast_fp16 = slice_by_index(begin = var_9039_begin_0, end = var_9039_end_0, end_mask = var_9039_end_mask_0, stride = var_9039_stride_0, x = q_51_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [4]> var_9049_begin_0 = const()[name = string("op_9049_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9049_end_0 = const()[name = string("op_9049_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9049_end_mask_0 = const()[name = string("op_9049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9049_stride_0 = const()[name = string("op_9049_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = slice_by_index(begin = var_9049_begin_0, end = var_9049_end_0, end_mask = var_9049_end_mask_0, stride = var_9049_stride_0, x = q_51_cast_fp16)[name = string("op_9049_cast_fp16")];
            int32 var_9051 = const()[name = string("op_9051"), val = int32(2)];
            bool stacked_51_interleave_0 = const()[name = string("stacked_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_51_cast_fp16 = concat(axis = var_9051, interleave = stacked_51_interleave_0, values = (var_9039_cast_fp16, var_9049_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9057_begin_0 = const()[name = string("op_9057_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9057_end_0 = const()[name = string("op_9057_end_0"), val = tensor<int32, [4]>([26, 8, 2064, 128])];
            tensor<bool, [4]> var_9057_end_mask_0 = const()[name = string("op_9057_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9057_cast_fp16 = slice_by_index(begin = var_9057_begin_0, end = var_9057_end_0, end_mask = var_9057_end_mask_0, x = read_state_0)[name = string("op_9057_cast_fp16")];
            tensor<int32, [4]> var_9067_begin_0 = const()[name = string("op_9067_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9067_end_0 = const()[name = string("op_9067_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9067_end_mask_0 = const()[name = string("op_9067_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9067_cast_fp16 = slice_by_index(begin = var_9067_begin_0, end = var_9067_end_0, end_mask = var_9067_end_mask_0, x = var_9057_cast_fp16)[name = string("op_9067_cast_fp16")];
            bool var_9076_transpose_x_1 = const()[name = string("op_9076_transpose_x_1"), val = bool(false)];
            bool var_9076_transpose_y_1 = const()[name = string("op_9076_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9076_cast_fp16 = matmul(transpose_x = var_9076_transpose_x_1, transpose_y = var_9076_transpose_y_1, x = var_9067_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<int32, [4]> var_9079_perm_0 = const()[name = string("op_9079_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9080_to_fp16 = const()[name = string("op_9080_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9079_cast_fp16 = transpose(perm = var_9079_perm_0, x = var_9076_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 2048]> var_9081_cast_fp16 = mul(x = var_9079_cast_fp16, y = var_9080_to_fp16)[name = string("op_9081_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_51_cast_fp16 = add(x = var_9081_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9087_transpose_x_1 = const()[name = string("op_9087_transpose_x_1"), val = bool(false)];
            bool var_9087_transpose_y_1 = const()[name = string("op_9087_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9087_cast_fp16 = matmul(transpose_x = var_9087_transpose_x_1, transpose_y = var_9087_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9089_cast_fp16 = mul(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("op_9089_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9089_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9093 = const()[name = string("op_9093"), val = int32(-1)];
            bool var_9094_interleave_0 = const()[name = string("op_9094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9094_cast_fp16 = concat(axis = var_9093, interleave = var_9094_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9094_cast_fp16")];
            int32 var_9095 = const()[name = string("op_9095"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_51_cast_fp16 = softmax(axis = var_9095, x = var_9094_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9102_begin_0 = const()[name = string("op_9102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9102_end_0 = const()[name = string("op_9102_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9102_end_mask_0 = const()[name = string("op_9102_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9102_cast_fp16 = slice_by_index(begin = var_9102_begin_0, end = var_9102_end_0, end_mask = var_9102_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9102_cast_fp16")];
            tensor<int32, [4]> var_9107_begin_0 = const()[name = string("op_9107_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9107_end_0 = const()[name = string("op_9107_end_0"), val = tensor<int32, [4]>([26, 8, 2064, 128])];
            tensor<bool, [4]> var_9107_end_mask_0 = const()[name = string("op_9107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9107_cast_fp16 = slice_by_index(begin = var_9107_begin_0, end = var_9107_end_0, end_mask = var_9107_end_mask_0, x = read_state_1)[name = string("op_9107_cast_fp16")];
            tensor<int32, [4]> var_9117_begin_0 = const()[name = string("op_9117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9117_end_0 = const()[name = string("op_9117_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9117_end_mask_0 = const()[name = string("op_9117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9117_cast_fp16 = slice_by_index(begin = var_9117_begin_0, end = var_9117_end_0, end_mask = var_9117_end_mask_0, x = var_9107_cast_fp16)[name = string("op_9117_cast_fp16")];
            bool var_9123_transpose_x_0 = const()[name = string("op_9123_transpose_x_0"), val = bool(false)];
            bool var_9123_transpose_y_0 = const()[name = string("op_9123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9123_cast_fp16 = matmul(transpose_x = var_9123_transpose_x_0, transpose_y = var_9123_transpose_y_0, x = var_9102_cast_fp16, y = var_9117_cast_fp16)[name = string("op_9123_cast_fp16")];
            tensor<int32, [4]> var_9128_begin_0 = const()[name = string("op_9128_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9128_end_0 = const()[name = string("op_9128_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9128_end_mask_0 = const()[name = string("op_9128_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9128_cast_fp16 = slice_by_index(begin = var_9128_begin_0, end = var_9128_end_0, end_mask = var_9128_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9128_cast_fp16")];
            bool var_9129_transpose_x_0 = const()[name = string("op_9129_transpose_x_0"), val = bool(false)];
            bool var_9129_transpose_y_0 = const()[name = string("op_9129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_8984_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9129_cast_fp16 = matmul(transpose_x = var_9129_transpose_x_0, transpose_y = var_9129_transpose_y_0, x = var_9128_cast_fp16, y = v_51_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9123_cast_fp16, y = var_9129_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9146_begin_0 = const()[name = string("op_9146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9146_end_0 = const()[name = string("op_9146_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9146_end_mask_0 = const()[name = string("op_9146_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9146_cast_fp16 = slice_by_index(begin = var_9146_begin_0, end = var_9146_end_0, end_mask = var_9146_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9146_cast_fp16")];
            tensor<int32, [4]> var_9161_begin_0 = const()[name = string("op_9161_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9161_end_0 = const()[name = string("op_9161_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9161_end_mask_0 = const()[name = string("op_9161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9161_cast_fp16 = slice_by_index(begin = var_9161_begin_0, end = var_9161_end_0, end_mask = var_9161_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9161_cast_fp16")];
            int32 var_9163 = const()[name = string("op_9163"), val = int32(1)];
            bool mixed_103_interleave_0 = const()[name = string("mixed_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_103_cast_fp16 = concat(axis = var_9163, interleave = mixed_103_interleave_0, values = (var_9146_cast_fp16, var_9161_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9167_perm_0 = const()[name = string("op_9167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9171 = const()[name = string("op_9171"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9167_cast_fp16 = transpose(perm = var_9167_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9171, x = var_9167_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_25_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709428672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711525888))))[name = string("model_layers_25_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_178_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_o_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_311_cast_fp16 = add(x = x_301_cast_fp16, y = linear_178_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_9182_promoted_to_fp16 = const()[name = string("op_9182_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9184_cast_fp16 = pow(x = x_311_cast_fp16, y = var_9182_promoted_to_fp16)[name = string("op_9184_cast_fp16")];
            tensor<int32, [1]> var_9186_axes_0 = const()[name = string("op_9186_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9186_keep_dims_0 = const()[name = string("op_9186_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9186_cast_fp16 = reduce_mean(axes = var_9186_axes_0, keep_dims = var_9186_keep_dims_0, x = var_9184_cast_fp16)[name = string("op_9186_cast_fp16")];
            fp16 var_9187_to_fp16 = const()[name = string("op_9187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9188_cast_fp16 = add(x = var_9186_cast_fp16, y = var_9187_to_fp16)[name = string("op_9188_cast_fp16")];
            fp32 var_9189_epsilon_0 = const()[name = string("op_9189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9189_cast_fp16 = rsqrt(epsilon = var_9189_epsilon_0, x = var_9188_cast_fp16)[name = string("op_9189_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9190_cast_fp16 = mul(x = x_311_cast_fp16, y = var_9189_cast_fp16)[name = string("op_9190_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711528000)))];
            tensor<fp16, [1, 16, 1024]> input_255_cast_fp16 = mul(x = var_9190_cast_fp16, y = model_layers_25_post_attention_layernorm_weight_to_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711530112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714675904))))[name = string("model_layers_25_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_gate_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9195_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714682112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717827904))))[name = string("model_layers_25_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_up_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_259_cast_fp16 = mul(x = var_9195_cast_fp16, y = linear_180_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_25_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717834112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720979904))))[name = string("model_layers_25_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_313_cast_fp16 = add(x = x_311_cast_fp16, y = linear_181_cast_fp16)[name = string("x_313_cast_fp16")];
            fp16 var_9209_promoted_to_fp16 = const()[name = string("op_9209_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9211_cast_fp16 = pow(x = x_313_cast_fp16, y = var_9209_promoted_to_fp16)[name = string("op_9211_cast_fp16")];
            tensor<int32, [1]> var_9213_axes_0 = const()[name = string("op_9213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9213_keep_dims_0 = const()[name = string("op_9213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9213_cast_fp16 = reduce_mean(axes = var_9213_axes_0, keep_dims = var_9213_keep_dims_0, x = var_9211_cast_fp16)[name = string("op_9213_cast_fp16")];
            fp16 var_9214_to_fp16 = const()[name = string("op_9214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9215_cast_fp16 = add(x = var_9213_cast_fp16, y = var_9214_to_fp16)[name = string("op_9215_cast_fp16")];
            fp32 var_9216_epsilon_0 = const()[name = string("op_9216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9216_cast_fp16 = rsqrt(epsilon = var_9216_epsilon_0, x = var_9215_cast_fp16)[name = string("op_9216_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9217_cast_fp16 = mul(x = x_313_cast_fp16, y = var_9216_cast_fp16)[name = string("op_9217_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720982016)))];
            tensor<fp16, [1, 16, 1024]> input_261_cast_fp16 = mul(x = var_9217_cast_fp16, y = model_layers_26_input_layernorm_weight_to_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_26_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720984128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723081344))))[name = string("model_layers_26_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_26_self_attn_q_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_9226 = const()[name = string("op_9226"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9227_cast_fp16 = reshape(shape = var_9226, x = linear_182_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<int32, [4]> x_315_perm_0 = const()[name = string("x_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9235_promoted_to_fp16 = const()[name = string("op_9235_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_315_cast_fp16 = transpose(perm = x_315_perm_0, x = var_9227_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9237_cast_fp16 = pow(x = x_315_cast_fp16, y = var_9235_promoted_to_fp16)[name = string("op_9237_cast_fp16")];
            tensor<int32, [1]> var_9239_axes_0 = const()[name = string("op_9239_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9239_keep_dims_0 = const()[name = string("op_9239_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9239_cast_fp16 = reduce_mean(axes = var_9239_axes_0, keep_dims = var_9239_keep_dims_0, x = var_9237_cast_fp16)[name = string("op_9239_cast_fp16")];
            fp16 var_9240_to_fp16 = const()[name = string("op_9240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9241_cast_fp16 = add(x = var_9239_cast_fp16, y = var_9240_to_fp16)[name = string("op_9241_cast_fp16")];
            fp32 var_9242_epsilon_0 = const()[name = string("op_9242_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9242_cast_fp16 = rsqrt(epsilon = var_9242_epsilon_0, x = var_9241_cast_fp16)[name = string("op_9242_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9243_cast_fp16 = mul(x = x_315_cast_fp16, y = var_9242_cast_fp16)[name = string("op_9243_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085504)))];
            tensor<fp16, [1, 16, 16, 128]> x_319_cast_fp16 = mul(x = var_9243_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_319_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724134464))))[name = string("model_layers_26_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_183_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_k_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_9252 = const()[name = string("op_9252"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9253_cast_fp16 = reshape(shape = var_9252, x = linear_183_cast_fp16)[name = string("op_9253_cast_fp16")];
            tensor<int32, [4]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9261_promoted_to_fp16 = const()[name = string("op_9261_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = var_9253_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_9263_cast_fp16 = pow(x = x_317_cast_fp16, y = var_9261_promoted_to_fp16)[name = string("op_9263_cast_fp16")];
            tensor<int32, [1]> var_9265_axes_0 = const()[name = string("op_9265_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9265_keep_dims_0 = const()[name = string("op_9265_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9265_cast_fp16 = reduce_mean(axes = var_9265_axes_0, keep_dims = var_9265_keep_dims_0, x = var_9263_cast_fp16)[name = string("op_9265_cast_fp16")];
            fp16 var_9266_to_fp16 = const()[name = string("op_9266_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9267_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9266_to_fp16)[name = string("op_9267_cast_fp16")];
            fp32 var_9268_epsilon_0 = const()[name = string("op_9268_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9268_cast_fp16 = rsqrt(epsilon = var_9268_epsilon_0, x = var_9267_cast_fp16)[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9269_cast_fp16 = mul(x = x_317_cast_fp16, y = var_9268_cast_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136576)))];
            tensor<fp16, [1, 8, 16, 128]> x_321_cast_fp16 = mul(x = var_9269_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_321_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725185536))))[name = string("model_layers_26_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_184_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_v_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_9278 = const()[name = string("op_9278"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9279_cast_fp16 = reshape(shape = var_9278, x = linear_184_cast_fp16)[name = string("op_9279_cast_fp16")];
            tensor<int32, [4]> v_53_perm_0 = const()[name = string("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_105_begin_0 = const()[name = string("first_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_105_end_0 = const()[name = string("first_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_105_end_mask_0 = const()[name = string("first_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_105_cast_fp16 = slice_by_index(begin = first_105_begin_0, end = first_105_end_0, end_mask = first_105_end_mask_0, x = x_319_cast_fp16)[name = string("first_105_cast_fp16")];
            tensor<int32, [4]> second_105_begin_0 = const()[name = string("second_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_105_end_0 = const()[name = string("second_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_105_end_mask_0 = const()[name = string("second_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_105_cast_fp16 = slice_by_index(begin = second_105_begin_0, end = second_105_end_0, end_mask = second_105_end_mask_0, x = x_319_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9293_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9294_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9296_cast_fp16 = sub(x = var_9293_cast_fp16, y = var_9294_cast_fp16)[name = string("op_9296_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9297_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9297_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9298_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9298_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9300_cast_fp16 = add(x = var_9297_cast_fp16, y = var_9298_cast_fp16)[name = string("op_9300_cast_fp16")];
            int32 var_9302 = const()[name = string("op_9302"), val = int32(-1)];
            bool q_53_interleave_0 = const()[name = string("q_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_53_cast_fp16 = concat(axis = var_9302, interleave = q_53_interleave_0, values = (var_9296_cast_fp16, var_9300_cast_fp16))[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> first_107_begin_0 = const()[name = string("first_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_107_end_0 = const()[name = string("first_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_107_end_mask_0 = const()[name = string("first_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_107_cast_fp16 = slice_by_index(begin = first_107_begin_0, end = first_107_end_0, end_mask = first_107_end_mask_0, x = x_321_cast_fp16)[name = string("first_107_cast_fp16")];
            tensor<int32, [4]> second_107_begin_0 = const()[name = string("second_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_107_end_0 = const()[name = string("second_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_107_end_mask_0 = const()[name = string("second_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_107_cast_fp16 = slice_by_index(begin = second_107_begin_0, end = second_107_end_0, end_mask = second_107_end_mask_0, x = x_321_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9314_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9315_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9317_cast_fp16 = sub(x = var_9314_cast_fp16, y = var_9315_cast_fp16)[name = string("op_9317_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9318_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9319_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9319_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9321_cast_fp16 = add(x = var_9318_cast_fp16, y = var_9319_cast_fp16)[name = string("op_9321_cast_fp16")];
            int32 var_9323 = const()[name = string("op_9323"), val = int32(-1)];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_53_cast_fp16 = concat(axis = var_9323, interleave = k_53_interleave_0, values = (var_9317_cast_fp16, var_9321_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_9334_begin_0 = const()[name = string("op_9334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9334_end_0 = const()[name = string("op_9334_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9334_end_mask_0 = const()[name = string("op_9334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9334_stride_0 = const()[name = string("op_9334_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9334_cast_fp16 = slice_by_index(begin = var_9334_begin_0, end = var_9334_end_0, end_mask = var_9334_end_mask_0, stride = var_9334_stride_0, x = q_53_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [4]> var_9344_begin_0 = const()[name = string("op_9344_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9344_end_0 = const()[name = string("op_9344_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9344_end_mask_0 = const()[name = string("op_9344_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9344_stride_0 = const()[name = string("op_9344_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9344_cast_fp16 = slice_by_index(begin = var_9344_begin_0, end = var_9344_end_0, end_mask = var_9344_end_mask_0, stride = var_9344_stride_0, x = q_53_cast_fp16)[name = string("op_9344_cast_fp16")];
            int32 var_9346 = const()[name = string("op_9346"), val = int32(2)];
            bool stacked_53_interleave_0 = const()[name = string("stacked_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_53_cast_fp16 = concat(axis = var_9346, interleave = stacked_53_interleave_0, values = (var_9334_cast_fp16, var_9344_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_9352_begin_0 = const()[name = string("op_9352_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9352_end_0 = const()[name = string("op_9352_end_0"), val = tensor<int32, [4]>([27, 8, 2064, 128])];
            tensor<bool, [4]> var_9352_end_mask_0 = const()[name = string("op_9352_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9352_cast_fp16 = slice_by_index(begin = var_9352_begin_0, end = var_9352_end_0, end_mask = var_9352_end_mask_0, x = read_state_0)[name = string("op_9352_cast_fp16")];
            tensor<int32, [4]> var_9362_begin_0 = const()[name = string("op_9362_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9362_end_0 = const()[name = string("op_9362_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9362_end_mask_0 = const()[name = string("op_9362_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9362_cast_fp16 = slice_by_index(begin = var_9362_begin_0, end = var_9362_end_0, end_mask = var_9362_end_mask_0, x = var_9352_cast_fp16)[name = string("op_9362_cast_fp16")];
            bool var_9371_transpose_x_1 = const()[name = string("op_9371_transpose_x_1"), val = bool(false)];
            bool var_9371_transpose_y_1 = const()[name = string("op_9371_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9371_cast_fp16 = matmul(transpose_x = var_9371_transpose_x_1, transpose_y = var_9371_transpose_y_1, x = var_9362_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<int32, [4]> var_9374_perm_0 = const()[name = string("op_9374_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9375_to_fp16 = const()[name = string("op_9375_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9374_cast_fp16 = transpose(perm = var_9374_perm_0, x = var_9371_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 2048]> var_9376_cast_fp16 = mul(x = var_9374_cast_fp16, y = var_9375_to_fp16)[name = string("op_9376_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_53_cast_fp16 = add(x = var_9376_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_9382_transpose_x_1 = const()[name = string("op_9382_transpose_x_1"), val = bool(false)];
            bool var_9382_transpose_y_1 = const()[name = string("op_9382_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9382_cast_fp16 = matmul(transpose_x = var_9382_transpose_x_1, transpose_y = var_9382_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_9382_cast_fp16")];
            fp16 var_9383_to_fp16 = const()[name = string("op_9383_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9384_cast_fp16 = mul(x = var_9382_cast_fp16, y = var_9383_to_fp16)[name = string("op_9384_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_9384_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_9388 = const()[name = string("op_9388"), val = int32(-1)];
            bool var_9389_interleave_0 = const()[name = string("op_9389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9389_cast_fp16 = concat(axis = var_9388, interleave = var_9389_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_9389_cast_fp16")];
            int32 var_9390 = const()[name = string("op_9390"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_53_cast_fp16 = softmax(axis = var_9390, x = var_9389_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_9397_begin_0 = const()[name = string("op_9397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9397_end_0 = const()[name = string("op_9397_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9397_end_mask_0 = const()[name = string("op_9397_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9397_cast_fp16 = slice_by_index(begin = var_9397_begin_0, end = var_9397_end_0, end_mask = var_9397_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<int32, [4]> var_9402_begin_0 = const()[name = string("op_9402_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9402_end_0 = const()[name = string("op_9402_end_0"), val = tensor<int32, [4]>([27, 8, 2064, 128])];
            tensor<bool, [4]> var_9402_end_mask_0 = const()[name = string("op_9402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9402_cast_fp16 = slice_by_index(begin = var_9402_begin_0, end = var_9402_end_0, end_mask = var_9402_end_mask_0, x = read_state_1)[name = string("op_9402_cast_fp16")];
            tensor<int32, [4]> var_9412_begin_0 = const()[name = string("op_9412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9412_end_0 = const()[name = string("op_9412_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9412_end_mask_0 = const()[name = string("op_9412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9412_cast_fp16 = slice_by_index(begin = var_9412_begin_0, end = var_9412_end_0, end_mask = var_9412_end_mask_0, x = var_9402_cast_fp16)[name = string("op_9412_cast_fp16")];
            bool var_9418_transpose_x_0 = const()[name = string("op_9418_transpose_x_0"), val = bool(false)];
            bool var_9418_transpose_y_0 = const()[name = string("op_9418_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9418_cast_fp16 = matmul(transpose_x = var_9418_transpose_x_0, transpose_y = var_9418_transpose_y_0, x = var_9397_cast_fp16, y = var_9412_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<int32, [4]> var_9423_begin_0 = const()[name = string("op_9423_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9423_end_0 = const()[name = string("op_9423_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9423_end_mask_0 = const()[name = string("op_9423_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9423_cast_fp16 = slice_by_index(begin = var_9423_begin_0, end = var_9423_end_0, end_mask = var_9423_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9423_cast_fp16")];
            bool var_9424_transpose_x_0 = const()[name = string("op_9424_transpose_x_0"), val = bool(false)];
            bool var_9424_transpose_y_0 = const()[name = string("op_9424_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_9279_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_9424_cast_fp16 = matmul(transpose_x = var_9424_transpose_x_0, transpose_y = var_9424_transpose_y_0, x = var_9423_cast_fp16, y = v_53_cast_fp16)[name = string("op_9424_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_9418_cast_fp16, y = var_9424_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_9441_begin_0 = const()[name = string("op_9441_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9441_end_0 = const()[name = string("op_9441_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9441_end_mask_0 = const()[name = string("op_9441_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9441_cast_fp16 = slice_by_index(begin = var_9441_begin_0, end = var_9441_end_0, end_mask = var_9441_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9441_cast_fp16")];
            tensor<int32, [4]> var_9456_begin_0 = const()[name = string("op_9456_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9456_end_0 = const()[name = string("op_9456_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9456_end_mask_0 = const()[name = string("op_9456_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9456_cast_fp16 = slice_by_index(begin = var_9456_begin_0, end = var_9456_end_0, end_mask = var_9456_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9456_cast_fp16")];
            int32 var_9458 = const()[name = string("op_9458"), val = int32(1)];
            bool mixed_107_interleave_0 = const()[name = string("mixed_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_107_cast_fp16 = concat(axis = var_9458, interleave = mixed_107_interleave_0, values = (var_9441_cast_fp16, var_9456_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_9462_perm_0 = const()[name = string("op_9462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9466 = const()[name = string("op_9466"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9462_cast_fp16 = transpose(perm = var_9462_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_9466, x = var_9462_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_26_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725187648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727284864))))[name = string("model_layers_26_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_o_proj_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = add(x = x_313_cast_fp16, y = linear_185_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_9477_promoted_to_fp16 = const()[name = string("op_9477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9479_cast_fp16 = pow(x = x_323_cast_fp16, y = var_9477_promoted_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<int32, [1]> var_9481_axes_0 = const()[name = string("op_9481_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9481_keep_dims_0 = const()[name = string("op_9481_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9481_cast_fp16 = reduce_mean(axes = var_9481_axes_0, keep_dims = var_9481_keep_dims_0, x = var_9479_cast_fp16)[name = string("op_9481_cast_fp16")];
            fp16 var_9482_to_fp16 = const()[name = string("op_9482_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9483_cast_fp16 = add(x = var_9481_cast_fp16, y = var_9482_to_fp16)[name = string("op_9483_cast_fp16")];
            fp32 var_9484_epsilon_0 = const()[name = string("op_9484_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9484_cast_fp16 = rsqrt(epsilon = var_9484_epsilon_0, x = var_9483_cast_fp16)[name = string("op_9484_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9485_cast_fp16 = mul(x = x_323_cast_fp16, y = var_9484_cast_fp16)[name = string("op_9485_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727286976)))];
            tensor<fp16, [1, 16, 1024]> input_265_cast_fp16 = mul(x = var_9485_cast_fp16, y = model_layers_26_post_attention_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727289088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730434880))))[name = string("model_layers_26_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_gate_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9490_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_9490_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730441088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733586880))))[name = string("model_layers_26_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_up_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_269_cast_fp16 = mul(x = var_9490_cast_fp16, y = linear_187_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_26_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733593088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736738880))))[name = string("model_layers_26_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_mlp_down_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_325_cast_fp16 = add(x = x_323_cast_fp16, y = linear_188_cast_fp16)[name = string("x_325_cast_fp16")];
            fp16 var_9504_promoted_to_fp16 = const()[name = string("op_9504_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9506_cast_fp16 = pow(x = x_325_cast_fp16, y = var_9504_promoted_to_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9508_axes_0 = const()[name = string("op_9508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9508_keep_dims_0 = const()[name = string("op_9508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9508_cast_fp16 = reduce_mean(axes = var_9508_axes_0, keep_dims = var_9508_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9508_cast_fp16")];
            fp16 var_9509_to_fp16 = const()[name = string("op_9509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9510_cast_fp16 = add(x = var_9508_cast_fp16, y = var_9509_to_fp16)[name = string("op_9510_cast_fp16")];
            fp32 var_9511_epsilon_0 = const()[name = string("op_9511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9511_cast_fp16 = rsqrt(epsilon = var_9511_epsilon_0, x = var_9510_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9512_cast_fp16 = mul(x = x_325_cast_fp16, y = var_9511_cast_fp16)[name = string("op_9512_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736740992)))];
            tensor<fp16, [1, 16, 1024]> input_271_cast_fp16 = mul(x = var_9512_cast_fp16, y = model_layers_27_input_layernorm_weight_to_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_27_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736743104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738840320))))[name = string("model_layers_27_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_27_self_attn_q_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<int32, [4]> var_9521 = const()[name = string("op_9521"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9522_cast_fp16 = reshape(shape = var_9521, x = linear_189_cast_fp16)[name = string("op_9522_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9530_promoted_to_fp16 = const()[name = string("op_9530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_9522_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_9532_cast_fp16 = pow(x = x_327_cast_fp16, y = var_9530_promoted_to_fp16)[name = string("op_9532_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9532_cast_fp16)[name = string("op_9534_cast_fp16")];
            fp16 var_9535_to_fp16 = const()[name = string("op_9535_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9536_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9535_to_fp16)[name = string("op_9536_cast_fp16")];
            fp32 var_9537_epsilon_0 = const()[name = string("op_9537_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9537_cast_fp16 = rsqrt(epsilon = var_9537_epsilon_0, x = var_9536_cast_fp16)[name = string("op_9537_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9538_cast_fp16 = mul(x = x_327_cast_fp16, y = var_9537_cast_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844480)))];
            tensor<fp16, [1, 16, 16, 128]> x_331_cast_fp16 = mul(x = var_9538_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_331_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739893440))))[name = string("model_layers_27_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_k_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9547 = const()[name = string("op_9547"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9548_cast_fp16 = reshape(shape = var_9547, x = linear_190_cast_fp16)[name = string("op_9548_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9556_promoted_to_fp16 = const()[name = string("op_9556_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_9548_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_9558_cast_fp16 = pow(x = x_329_cast_fp16, y = var_9556_promoted_to_fp16)[name = string("op_9558_cast_fp16")];
            tensor<int32, [1]> var_9560_axes_0 = const()[name = string("op_9560_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9560_keep_dims_0 = const()[name = string("op_9560_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9560_cast_fp16 = reduce_mean(axes = var_9560_axes_0, keep_dims = var_9560_keep_dims_0, x = var_9558_cast_fp16)[name = string("op_9560_cast_fp16")];
            fp16 var_9561_to_fp16 = const()[name = string("op_9561_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9562_cast_fp16 = add(x = var_9560_cast_fp16, y = var_9561_to_fp16)[name = string("op_9562_cast_fp16")];
            fp32 var_9563_epsilon_0 = const()[name = string("op_9563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9563_cast_fp16 = rsqrt(epsilon = var_9563_epsilon_0, x = var_9562_cast_fp16)[name = string("op_9563_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9564_cast_fp16 = mul(x = x_329_cast_fp16, y = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895552)))];
            tensor<fp16, [1, 8, 16, 128]> x_333_cast_fp16 = mul(x = var_9564_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_333_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740944512))))[name = string("model_layers_27_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_191_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_v_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9573 = const()[name = string("op_9573"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9574_cast_fp16 = reshape(shape = var_9573, x = linear_191_cast_fp16)[name = string("op_9574_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_109_begin_0 = const()[name = string("first_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_109_end_0 = const()[name = string("first_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_109_end_mask_0 = const()[name = string("first_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_109_cast_fp16 = slice_by_index(begin = first_109_begin_0, end = first_109_end_0, end_mask = first_109_end_mask_0, x = x_331_cast_fp16)[name = string("first_109_cast_fp16")];
            tensor<int32, [4]> second_109_begin_0 = const()[name = string("second_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_109_end_0 = const()[name = string("second_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_109_end_mask_0 = const()[name = string("second_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_109_cast_fp16 = slice_by_index(begin = second_109_begin_0, end = second_109_end_0, end_mask = second_109_end_mask_0, x = x_331_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9588_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9588_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9589_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9589_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9591_cast_fp16 = sub(x = var_9588_cast_fp16, y = var_9589_cast_fp16)[name = string("op_9591_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9592_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9593_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9593_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9595_cast_fp16 = add(x = var_9592_cast_fp16, y = var_9593_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_cast_fp16 = concat(axis = var_9597, interleave = q_interleave_0, values = (var_9591_cast_fp16, var_9595_cast_fp16))[name = string("q_cast_fp16")];
            tensor<int32, [4]> first_begin_0 = const()[name = string("first_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_end_0 = const()[name = string("first_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_end_mask_0 = const()[name = string("first_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_cast_fp16 = slice_by_index(begin = first_begin_0, end = first_end_0, end_mask = first_end_mask_0, x = x_333_cast_fp16)[name = string("first_cast_fp16")];
            tensor<int32, [4]> second_begin_0 = const()[name = string("second_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_end_0 = const()[name = string("second_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_end_mask_0 = const()[name = string("second_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_cast_fp16 = slice_by_index(begin = second_begin_0, end = second_end_0, end_mask = second_end_mask_0, x = x_333_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9609_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9609_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9610_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9610_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9612_cast_fp16 = sub(x = var_9609_cast_fp16, y = var_9610_cast_fp16)[name = string("op_9612_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9613_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9614_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9614_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9616_cast_fp16 = add(x = var_9613_cast_fp16, y = var_9614_cast_fp16)[name = string("op_9616_cast_fp16")];
            int32 var_9618 = const()[name = string("op_9618"), val = int32(-1)];
            bool k_interleave_0 = const()[name = string("k_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_cast_fp16 = concat(axis = var_9618, interleave = k_interleave_0, values = (var_9612_cast_fp16, var_9616_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_9629_begin_0 = const()[name = string("op_9629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9629_end_0 = const()[name = string("op_9629_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9629_end_mask_0 = const()[name = string("op_9629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9629_stride_0 = const()[name = string("op_9629_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9629_cast_fp16 = slice_by_index(begin = var_9629_begin_0, end = var_9629_end_0, end_mask = var_9629_end_mask_0, stride = var_9629_stride_0, x = q_cast_fp16)[name = string("op_9629_cast_fp16")];
            tensor<int32, [4]> var_9639_begin_0 = const()[name = string("op_9639_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9639_end_0 = const()[name = string("op_9639_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9639_end_mask_0 = const()[name = string("op_9639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9639_stride_0 = const()[name = string("op_9639_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9639_cast_fp16 = slice_by_index(begin = var_9639_begin_0, end = var_9639_end_0, end_mask = var_9639_end_mask_0, stride = var_9639_stride_0, x = q_cast_fp16)[name = string("op_9639_cast_fp16")];
            int32 var_9641 = const()[name = string("op_9641"), val = int32(2)];
            bool stacked_interleave_0 = const()[name = string("stacked_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_cast_fp16 = concat(axis = var_9641, interleave = stacked_interleave_0, values = (var_9629_cast_fp16, var_9639_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_9647_begin_0 = const()[name = string("op_9647_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9647_end_0 = const()[name = string("op_9647_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_9647_end_mask_0 = const()[name = string("op_9647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9647_cast_fp16 = slice_by_index(begin = var_9647_begin_0, end = var_9647_end_0, end_mask = var_9647_end_mask_0, x = read_state_0)[name = string("op_9647_cast_fp16")];
            tensor<int32, [4]> var_9657_begin_0 = const()[name = string("op_9657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9657_end_0 = const()[name = string("op_9657_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9657_end_mask_0 = const()[name = string("op_9657_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9657_cast_fp16 = slice_by_index(begin = var_9657_begin_0, end = var_9657_end_0, end_mask = var_9657_end_mask_0, x = var_9647_cast_fp16)[name = string("op_9657_cast_fp16")];
            bool var_9666_transpose_x_1 = const()[name = string("op_9666_transpose_x_1"), val = bool(false)];
            bool var_9666_transpose_y_1 = const()[name = string("op_9666_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 2048, 32]> var_9666_cast_fp16 = matmul(transpose_x = var_9666_transpose_x_1, transpose_y = var_9666_transpose_y_1, x = var_9657_cast_fp16, y = stacked_cast_fp16)[name = string("op_9666_cast_fp16")];
            tensor<int32, [4]> var_9669_perm_0 = const()[name = string("op_9669_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9670_to_fp16 = const()[name = string("op_9670_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 2048]> var_9669_cast_fp16 = transpose(perm = var_9669_perm_0, x = var_9666_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 2048]> var_9671_cast_fp16 = mul(x = var_9669_cast_fp16, y = var_9670_to_fp16)[name = string("op_9671_cast_fp16")];
            tensor<fp16, [1, 8, 32, 2048]> cache_weights_cast_fp16 = add(x = var_9671_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_9677_transpose_x_1 = const()[name = string("op_9677_transpose_x_1"), val = bool(false)];
            bool var_9677_transpose_y_1 = const()[name = string("op_9677_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9677_cast_fp16 = matmul(transpose_x = var_9677_transpose_x_1, transpose_y = var_9677_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_9677_cast_fp16")];
            fp16 var_9678_to_fp16 = const()[name = string("op_9678_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9679_cast_fp16 = mul(x = var_9677_cast_fp16, y = var_9678_to_fp16)[name = string("op_9679_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_9679_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_9683 = const()[name = string("op_9683"), val = int32(-1)];
            bool var_9684_interleave_0 = const()[name = string("op_9684_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 2064]> var_9684_cast_fp16 = concat(axis = var_9683, interleave = var_9684_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_9684_cast_fp16")];
            int32 var_9685 = const()[name = string("op_9685"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 2064]> shares_cast_fp16 = softmax(axis = var_9685, x = var_9684_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_9692_begin_0 = const()[name = string("op_9692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9692_end_0 = const()[name = string("op_9692_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2048])];
            tensor<bool, [4]> var_9692_end_mask_0 = const()[name = string("op_9692_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 2048]> var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, x = shares_cast_fp16)[name = string("op_9692_cast_fp16")];
            tensor<int32, [4]> var_9697_begin_0 = const()[name = string("op_9697_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9697_end_0 = const()[name = string("op_9697_end_0"), val = tensor<int32, [4]>([1, 8, 2064, 128])];
            tensor<bool, [4]> var_9697_end_mask_0 = const()[name = string("op_9697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 2064, 128]> var_9697_cast_fp16 = slice_by_index(begin = var_9697_begin_0, end = var_9697_end_0, end_mask = var_9697_end_mask_0, x = read_state_1)[name = string("op_9697_cast_fp16")];
            tensor<int32, [4]> var_9707_begin_0 = const()[name = string("op_9707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9707_end_0 = const()[name = string("op_9707_end_0"), val = tensor<int32, [4]>([1, 8, 2048, 128])];
            tensor<bool, [4]> var_9707_end_mask_0 = const()[name = string("op_9707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 2048, 128]> var_9707_cast_fp16 = slice_by_index(begin = var_9707_begin_0, end = var_9707_end_0, end_mask = var_9707_end_mask_0, x = var_9697_cast_fp16)[name = string("op_9707_cast_fp16")];
            bool var_9713_transpose_x_0 = const()[name = string("op_9713_transpose_x_0"), val = bool(false)];
            bool var_9713_transpose_y_0 = const()[name = string("op_9713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9713_cast_fp16 = matmul(transpose_x = var_9713_transpose_x_0, transpose_y = var_9713_transpose_y_0, x = var_9692_cast_fp16, y = var_9707_cast_fp16)[name = string("op_9713_cast_fp16")];
            tensor<int32, [4]> var_9718_begin_0 = const()[name = string("op_9718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2048])];
            tensor<int32, [4]> var_9718_end_0 = const()[name = string("op_9718_end_0"), val = tensor<int32, [4]>([1, 8, 32, 2064])];
            tensor<bool, [4]> var_9718_end_mask_0 = const()[name = string("op_9718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9718_cast_fp16 = slice_by_index(begin = var_9718_begin_0, end = var_9718_end_0, end_mask = var_9718_end_mask_0, x = shares_cast_fp16)[name = string("op_9718_cast_fp16")];
            bool var_9719_transpose_x_0 = const()[name = string("op_9719_transpose_x_0"), val = bool(false)];
            bool var_9719_transpose_y_0 = const()[name = string("op_9719_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_9574_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_9719_cast_fp16 = matmul(transpose_x = var_9719_transpose_x_0, transpose_y = var_9719_transpose_y_0, x = var_9718_cast_fp16, y = v_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_9713_cast_fp16, y = var_9719_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_9736_begin_0 = const()[name = string("op_9736_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9736_end_0 = const()[name = string("op_9736_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9736_end_mask_0 = const()[name = string("op_9736_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9736_cast_fp16 = slice_by_index(begin = var_9736_begin_0, end = var_9736_end_0, end_mask = var_9736_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9736_cast_fp16")];
            tensor<int32, [4]> var_9751_begin_0 = const()[name = string("op_9751_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9751_end_0 = const()[name = string("op_9751_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9751_end_mask_0 = const()[name = string("op_9751_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9751_cast_fp16 = slice_by_index(begin = var_9751_begin_0, end = var_9751_end_0, end_mask = var_9751_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9751_cast_fp16")];
            int32 var_9753 = const()[name = string("op_9753"), val = int32(1)];
            bool mixed_interleave_0 = const()[name = string("mixed_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_cast_fp16 = concat(axis = var_9753, interleave = mixed_interleave_0, values = (var_9736_cast_fp16, var_9751_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_9757_perm_0 = const()[name = string("op_9757_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9761 = const()[name = string("op_9761"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9757_cast_fp16 = transpose(perm = var_9757_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_9761, x = var_9757_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_27_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740946624))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743043840))))[name = string("model_layers_27_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_192_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_o_proj_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = linear_192_cast_fp16)[name = string("x_335_cast_fp16")];
            fp16 var_9772_promoted_to_fp16 = const()[name = string("op_9772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9774_cast_fp16 = pow(x = x_335_cast_fp16, y = var_9772_promoted_to_fp16)[name = string("op_9774_cast_fp16")];
            tensor<int32, [1]> var_9776_axes_0 = const()[name = string("op_9776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9776_keep_dims_0 = const()[name = string("op_9776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9776_cast_fp16 = reduce_mean(axes = var_9776_axes_0, keep_dims = var_9776_keep_dims_0, x = var_9774_cast_fp16)[name = string("op_9776_cast_fp16")];
            fp16 var_9777_to_fp16 = const()[name = string("op_9777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9778_cast_fp16 = add(x = var_9776_cast_fp16, y = var_9777_to_fp16)[name = string("op_9778_cast_fp16")];
            fp32 var_9779_epsilon_0 = const()[name = string("op_9779_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9779_cast_fp16 = rsqrt(epsilon = var_9779_epsilon_0, x = var_9778_cast_fp16)[name = string("op_9779_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9780_cast_fp16 = mul(x = x_335_cast_fp16, y = var_9779_cast_fp16)[name = string("op_9780_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743045952)))];
            tensor<fp16, [1, 16, 1024]> input_275_cast_fp16 = mul(x = var_9780_cast_fp16, y = model_layers_27_post_attention_layernorm_weight_to_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743048064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746193856))))[name = string("model_layers_27_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_gate_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9785_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_9785_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746200064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749345856))))[name = string("model_layers_27_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_up_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_cast_fp16 = mul(x = var_9785_cast_fp16, y = linear_194_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_27_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749352064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752497856))))[name = string("model_layers_27_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_195_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_mlp_down_proj_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_337_cast_fp16 = add(x = x_335_cast_fp16, y = linear_195_cast_fp16)[name = string("x_337_cast_fp16")];
            int32 var_9795 = const()[name = string("op_9795"), val = int32(1)];
            int32 x_batch_dims_0 = const()[name = string("x_batch_dims_0"), val = int32(0)];
            bool x_validate_indices_0 = const()[name = string("x_validate_indices_0"), val = bool(false)];
            string answerRow_to_uint16_dtype_0 = const()[name = string("answerRow_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> answerRow_to_uint16 = cast(dtype = answerRow_to_uint16_dtype_0, x = answerRow)[name = string("cast_0")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16_cast_uint16 = gather(axis = var_9795, batch_dims = x_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_validate_indices_0, x = x_337_cast_fp16)[name = string("x_cast_fp16_cast_uint16")];
            fp16 var_9801_promoted_to_fp16 = const()[name = string("op_9801_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_9803_cast_fp16 = pow(x = x_cast_fp16_cast_uint16, y = var_9801_promoted_to_fp16)[name = string("op_9803_cast_fp16")];
            tensor<int32, [1]> var_9805_axes_0 = const()[name = string("op_9805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9805_keep_dims_0 = const()[name = string("op_9805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_9805_cast_fp16 = reduce_mean(axes = var_9805_axes_0, keep_dims = var_9805_keep_dims_0, x = var_9803_cast_fp16)[name = string("op_9805_cast_fp16")];
            fp16 var_9806_to_fp16 = const()[name = string("op_9806_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_9807_cast_fp16 = add(x = var_9805_cast_fp16, y = var_9806_to_fp16)[name = string("op_9807_cast_fp16")];
            fp32 var_9808_epsilon_0 = const()[name = string("op_9808_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_9808_cast_fp16 = rsqrt(epsilon = var_9808_epsilon_0, x = var_9807_cast_fp16)[name = string("op_9808_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_9809_cast_fp16 = mul(x = x_cast_fp16_cast_uint16, y = var_9808_cast_fp16)[name = string("op_9809_cast_fp16")];
            tensor<fp16, [1024]> model_norm_weight_to_fp16 = const()[name = string("model_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752499968)))];
            tensor<fp16, [1, 1, 1024]> last_cast_fp16 = mul(x = var_9809_cast_fp16, y = model_norm_weight_to_fp16)[name = string("last_cast_fp16")];
            tensor<fp16, [151936]> linear_196_bias_0_to_fp16 = const()[name = string("linear_196_bias_0_to_fp16"), val = tensor<fp16, [151936]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752502080)))];
            tensor<fp16, [1, 1, 151936]> linear_196_cast_fp16 = linear(bias = linear_196_bias_0_to_fp16, weight = model_embed_tokens_weight_to_fp16, x = last_cast_fp16)[name = string("linear_196_cast_fp16")];
            int32 var_9815_axis_0 = const()[name = string("op_9815_axis_0"), val = int32(-1)];
            bool var_9815_keep_dims_0 = const()[name = string("op_9815_keep_dims_0"), val = bool(false)];
            string var_9815_output_dtype_0 = const()[name = string("op_9815_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_9815_axis_0, keep_dims = var_9815_keep_dims_0, output_dtype = var_9815_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_9815_cast_fp16")];
            int32 var_9822 = const()[name = string("op_9822"), val = int32(0)];
            bool var_9823_interleave_0 = const()[name = string("op_9823_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_9822, interleave = var_9823_interleave_0, values = (k_1_cast_fp16, k_3_cast_fp16, k_5_cast_fp16, k_7_cast_fp16, k_9_cast_fp16, k_11_cast_fp16, k_13_cast_fp16, k_15_cast_fp16, k_17_cast_fp16, k_19_cast_fp16, k_21_cast_fp16, k_23_cast_fp16, k_25_cast_fp16, k_27_cast_fp16, k_29_cast_fp16, k_31_cast_fp16, k_33_cast_fp16, k_35_cast_fp16, k_37_cast_fp16, k_39_cast_fp16, k_41_cast_fp16, k_43_cast_fp16, k_45_cast_fp16, k_47_cast_fp16, k_49_cast_fp16, k_51_cast_fp16, k_53_cast_fp16, k_cast_fp16))[name = string("op_9823_cast_fp16")];
            int32 var_9825 = const()[name = string("op_9825"), val = int32(0)];
            bool var_9826_interleave_0 = const()[name = string("op_9826_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_9825, interleave = var_9826_interleave_0, values = (v_1_cast_fp16, v_3_cast_fp16, v_5_cast_fp16, v_7_cast_fp16, v_9_cast_fp16, v_11_cast_fp16, v_13_cast_fp16, v_15_cast_fp16, v_17_cast_fp16, v_19_cast_fp16, v_21_cast_fp16, v_23_cast_fp16, v_25_cast_fp16, v_27_cast_fp16, v_29_cast_fp16, v_31_cast_fp16, v_33_cast_fp16, v_35_cast_fp16, v_37_cast_fp16, v_39_cast_fp16, v_41_cast_fp16, v_43_cast_fp16, v_45_cast_fp16, v_47_cast_fp16, v_49_cast_fp16, v_51_cast_fp16, v_53_cast_fp16, v_cast_fp16))[name = string("op_9826_cast_fp16")];
        } -> (nextToken, newKeys, newValues);
    func window4096<ios18>(tensor<int32, [1]> answerRow, tensor<fp16, [1, 1, 1, 4096]> cacheMask, tensor<int32, [1, 16]> inputIds, state<tensor<fp16, [28, 8, 4112, 128]>> keyCache, tensor<int32, [16]> positions, state<tensor<fp16, [28, 8, 4112, 128]>> valueCache) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> model_embed_tokens_weight_to_fp16 = const()[name = string("model_embed_tokens_weight_to_fp16"), val = tensor<fp16, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = inputIds, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0 = add(x = inputIds, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = inputIds, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 16]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 x_1_cast_fp16_axis_0 = const()[name = string("x_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 16, 1024]> x_1_cast_fp16 = gather(axis = x_1_cast_fp16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = model_embed_tokens_weight_to_fp16)[name = string("x_1_cast_fp16")];
            int32 var_1504 = const()[name = string("op_1504"), val = int32(0)];
            int32 var_1505_batch_dims_0 = const()[name = string("op_1505_batch_dims_0"), val = int32(0)];
            bool var_1505_validate_indices_0 = const()[name = string("op_1505_validate_indices_0"), val = bool(false)];
            tensor<fp16, [4112, 64]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4112, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753078784))), scale = tensor<fp16, [4112, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753342016))))[name = string("cos_to_fp16_quantized")];
            string positions_to_uint16_dtype_0 = const()[name = string("positions_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [16]> positions_to_uint16 = cast(dtype = positions_to_uint16_dtype_0, x = positions)[name = string("cast_1")];
            tensor<fp16, [16, 64]> var_1505_cast_fp16_cast_uint16 = gather(axis = var_1504, batch_dims = var_1505_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1505_validate_indices_0, x = cos_to_fp16_quantized)[name = string("op_1505_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> cos_1_cast_fp16 = reshape(shape = var_1510, x = var_1505_cast_fp16_cast_uint16)[name = string("cos_1_cast_fp16")];
            int32 var_1512 = const()[name = string("op_1512"), val = int32(0)];
            int32 var_1513_batch_dims_0 = const()[name = string("op_1513_batch_dims_0"), val = int32(0)];
            bool var_1513_validate_indices_0 = const()[name = string("op_1513_validate_indices_0"), val = bool(false)];
            tensor<fp16, [4112, 64]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [4112, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753350336))), scale = tensor<fp16, [4112, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(753613568))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [16, 64]> var_1513_cast_fp16_cast_uint16 = gather(axis = var_1512, batch_dims = var_1513_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1513_validate_indices_0, x = sin_to_fp16_quantized)[name = string("op_1513_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1518 = const()[name = string("op_1518"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> sin_1_cast_fp16 = reshape(shape = var_1518, x = var_1513_cast_fp16_cast_uint16)[name = string("sin_1_cast_fp16")];
            fp16 var_1539_promoted_to_fp16 = const()[name = string("op_1539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = pow(x = x_1_cast_fp16, y = var_1539_promoted_to_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1541_cast_fp16)[name = string("op_1543_cast_fp16")];
            fp16 var_1544_to_fp16 = const()[name = string("op_1544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1545_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1544_to_fp16)[name = string("op_1545_cast_fp16")];
            fp32 var_1546_epsilon_0 = const()[name = string("op_1546_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1546_cast_fp16 = rsqrt(epsilon = var_1546_epsilon_0, x = var_1545_cast_fp16)[name = string("op_1546_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1547_cast_fp16 = mul(x = x_1_cast_fp16, y = var_1546_cast_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311235072)))];
            tensor<fp16, [1, 16, 1024]> input_1_cast_fp16 = mul(x = var_1547_cast_fp16, y = model_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311237184))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313334400))))[name = string("model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313338560)))];
            tensor<fp16, [1, 16, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_1556 = const()[name = string("op_1556"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1557_cast_fp16 = reshape(shape = var_1556, x = linear_0_cast_fp16)[name = string("op_1557_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1565_promoted_to_fp16 = const()[name = string("op_1565_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_1557_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1567_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1565_promoted_to_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [1]> var_1569_axes_0 = const()[name = string("op_1569_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1569_keep_dims_0 = const()[name = string("op_1569_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1569_cast_fp16 = reduce_mean(axes = var_1569_axes_0, keep_dims = var_1569_keep_dims_0, x = var_1567_cast_fp16)[name = string("op_1569_cast_fp16")];
            fp16 var_1570_to_fp16 = const()[name = string("op_1570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1571_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_to_fp16)[name = string("op_1571_cast_fp16")];
            fp32 var_1572_epsilon_0 = const()[name = string("op_1572_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1572_cast_fp16 = rsqrt(epsilon = var_1572_epsilon_0, x = var_1571_cast_fp16)[name = string("op_1572_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1573_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1572_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313342720)))];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = mul(x = var_1573_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313343040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314391680))))[name = string("model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314393792)))];
            tensor<fp16, [1, 16, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = string("op_1582"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1583_cast_fp16 = reshape(shape = var_1582, x = linear_1_cast_fp16)[name = string("op_1583_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1591_promoted_to_fp16 = const()[name = string("op_1591_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1583_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1593_cast_fp16 = pow(x = x_5_cast_fp16, y = var_1591_promoted_to_fp16)[name = string("op_1593_cast_fp16")];
            tensor<int32, [1]> var_1595_axes_0 = const()[name = string("op_1595_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1595_keep_dims_0 = const()[name = string("op_1595_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1595_cast_fp16 = reduce_mean(axes = var_1595_axes_0, keep_dims = var_1595_keep_dims_0, x = var_1593_cast_fp16)[name = string("op_1595_cast_fp16")];
            fp16 var_1596_to_fp16 = const()[name = string("op_1596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1597_cast_fp16 = add(x = var_1595_cast_fp16, y = var_1596_to_fp16)[name = string("op_1597_cast_fp16")];
            fp32 var_1598_epsilon_0 = const()[name = string("op_1598_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1598_cast_fp16 = rsqrt(epsilon = var_1598_epsilon_0, x = var_1597_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1599_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1598_cast_fp16)[name = string("op_1599_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314395904)))];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = mul(x = var_1599_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314396224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315444864))))[name = string("model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1608 = const()[name = string("op_1608"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1609_cast_fp16 = reshape(shape = var_1608, x = linear_2_cast_fp16)[name = string("op_1609_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_1_begin_0 = const()[name = string("first_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_1_end_0 = const()[name = string("first_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_1_end_mask_0 = const()[name = string("first_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_1_cast_fp16 = slice_by_index(begin = first_1_begin_0, end = first_1_end_0, end_mask = first_1_end_mask_0, x = x_7_cast_fp16)[name = string("first_1_cast_fp16")];
            tensor<int32, [4]> second_1_begin_0 = const()[name = string("second_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_1_end_0 = const()[name = string("second_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_1_end_mask_0 = const()[name = string("second_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_1_cast_fp16 = slice_by_index(begin = second_1_begin_0, end = second_1_end_0, end_mask = second_1_end_mask_0, x = x_7_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1623_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1624_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1626_cast_fp16 = sub(x = var_1623_cast_fp16, y = var_1624_cast_fp16)[name = string("op_1626_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1627_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1628_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1630_cast_fp16 = add(x = var_1627_cast_fp16, y = var_1628_cast_fp16)[name = string("op_1630_cast_fp16")];
            int32 var_1632 = const()[name = string("op_1632"), val = int32(-1)];
            bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_1_cast_fp16 = concat(axis = var_1632, interleave = q_1_interleave_0, values = (var_1626_cast_fp16, var_1630_cast_fp16))[name = string("q_1_cast_fp16")];
            tensor<int32, [4]> first_3_begin_0 = const()[name = string("first_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_3_end_0 = const()[name = string("first_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_3_end_mask_0 = const()[name = string("first_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_3_cast_fp16 = slice_by_index(begin = first_3_begin_0, end = first_3_end_0, end_mask = first_3_end_mask_0, x = x_9_cast_fp16)[name = string("first_3_cast_fp16")];
            tensor<int32, [4]> second_3_begin_0 = const()[name = string("second_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_3_end_0 = const()[name = string("second_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_3_end_mask_0 = const()[name = string("second_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_3_cast_fp16 = slice_by_index(begin = second_3_begin_0, end = second_3_end_0, end_mask = second_3_end_mask_0, x = x_9_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1644_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1645_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1647_cast_fp16 = sub(x = var_1644_cast_fp16, y = var_1645_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1648_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1649_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1651_cast_fp16 = add(x = var_1648_cast_fp16, y = var_1649_cast_fp16)[name = string("op_1651_cast_fp16")];
            int32 var_1653 = const()[name = string("op_1653"), val = int32(-1)];
            bool k_1_interleave_0 = const()[name = string("k_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_1_cast_fp16 = concat(axis = var_1653, interleave = k_1_interleave_0, values = (var_1647_cast_fp16, var_1651_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1664_begin_0 = const()[name = string("op_1664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1664_end_0 = const()[name = string("op_1664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1664_end_mask_0 = const()[name = string("op_1664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1664_stride_0 = const()[name = string("op_1664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, stride = var_1664_stride_0, x = q_1_cast_fp16)[name = string("op_1664_cast_fp16")];
            tensor<int32, [4]> var_1674_begin_0 = const()[name = string("op_1674_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1674_end_0 = const()[name = string("op_1674_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1674_end_mask_0 = const()[name = string("op_1674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1674_stride_0 = const()[name = string("op_1674_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1674_cast_fp16 = slice_by_index(begin = var_1674_begin_0, end = var_1674_end_0, end_mask = var_1674_end_mask_0, stride = var_1674_stride_0, x = q_1_cast_fp16)[name = string("op_1674_cast_fp16")];
            int32 var_1676 = const()[name = string("op_1676"), val = int32(2)];
            bool stacked_1_interleave_0 = const()[name = string("stacked_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_1_cast_fp16 = concat(axis = var_1676, interleave = stacked_1_interleave_0, values = (var_1664_cast_fp16, var_1674_cast_fp16))[name = string("stacked_1_cast_fp16")];
            tensor<fp16, [28, 8, 4112, 128]> read_state_0 = read_state(input = keyCache)[name = string("read_state_0")];
            tensor<int32, [4]> var_1682_begin_0 = const()[name = string("op_1682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1682_end_0 = const()[name = string("op_1682_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_1682_end_mask_0 = const()[name = string("op_1682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, x = read_state_0)[name = string("op_1682_cast_fp16")];
            tensor<int32, [4]> var_1692_begin_0 = const()[name = string("op_1692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1692_end_0 = const()[name = string("op_1692_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_1692_end_mask_0 = const()[name = string("op_1692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_1692_cast_fp16 = slice_by_index(begin = var_1692_begin_0, end = var_1692_end_0, end_mask = var_1692_end_mask_0, x = var_1682_cast_fp16)[name = string("op_1692_cast_fp16")];
            bool var_1701_transpose_x_1 = const()[name = string("op_1701_transpose_x_1"), val = bool(false)];
            bool var_1701_transpose_y_1 = const()[name = string("op_1701_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_1701_cast_fp16 = matmul(transpose_x = var_1701_transpose_x_1, transpose_y = var_1701_transpose_y_1, x = var_1692_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> var_1704_perm_0 = const()[name = string("op_1704_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_1704_cast_fp16 = transpose(perm = var_1704_perm_0, x = var_1701_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 4096]> var_1706_cast_fp16 = mul(x = var_1704_cast_fp16, y = var_1705_to_fp16)[name = string("op_1706_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_1_cast_fp16 = add(x = var_1706_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1712_transpose_x_1 = const()[name = string("op_1712_transpose_x_1"), val = bool(false)];
            bool var_1712_transpose_y_1 = const()[name = string("op_1712_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1712_cast_fp16 = matmul(transpose_x = var_1712_transpose_x_1, transpose_y = var_1712_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1712_cast_fp16")];
            fp16 var_1713_to_fp16 = const()[name = string("op_1713_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1714_cast_fp16 = mul(x = var_1712_cast_fp16, y = var_1713_to_fp16)[name = string("op_1714_cast_fp16")];
            tensor<fp16, [32, 16]> later_rows_mask_to_fp16 = const()[name = string("later_rows_mask_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315446976)))];
            tensor<fp16, [1, 8, 32, 16]> call_weights_1_cast_fp16 = add(x = var_1714_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_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, 8, 32, 4112]> var_1719_cast_fp16 = concat(axis = var_1718, interleave = var_1719_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1719_cast_fp16")];
            int32 var_1720 = const()[name = string("op_1720"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_1_cast_fp16 = softmax(axis = var_1720, x = var_1719_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1727_begin_0 = const()[name = string("op_1727_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1727_end_0 = const()[name = string("op_1727_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_1727_end_mask_0 = const()[name = string("op_1727_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_1727_cast_fp16 = slice_by_index(begin = var_1727_begin_0, end = var_1727_end_0, end_mask = var_1727_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [28, 8, 4112, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = read_state_1)[name = string("op_1732_cast_fp16")];
            tensor<int32, [4]> var_1742_begin_0 = const()[name = string("op_1742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1742_end_0 = const()[name = string("op_1742_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_1742_end_mask_0 = const()[name = string("op_1742_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_1742_cast_fp16 = slice_by_index(begin = var_1742_begin_0, end = var_1742_end_0, end_mask = var_1742_end_mask_0, x = var_1732_cast_fp16)[name = string("op_1742_cast_fp16")];
            bool var_1748_transpose_x_0 = const()[name = string("op_1748_transpose_x_0"), val = bool(false)];
            bool var_1748_transpose_y_0 = const()[name = string("op_1748_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1748_cast_fp16 = matmul(transpose_x = var_1748_transpose_x_0, transpose_y = var_1748_transpose_y_0, x = var_1727_cast_fp16, y = var_1742_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1753_cast_fp16 = slice_by_index(begin = var_1753_begin_0, end = var_1753_end_0, end_mask = var_1753_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1753_cast_fp16")];
            bool var_1754_transpose_x_0 = const()[name = string("op_1754_transpose_x_0"), val = bool(false)];
            bool var_1754_transpose_y_0 = const()[name = string("op_1754_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1609_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1754_cast_fp16 = matmul(transpose_x = var_1754_transpose_x_0, transpose_y = var_1754_transpose_y_0, x = var_1753_cast_fp16, y = v_1_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1754_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1771_begin_0 = const()[name = string("op_1771_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1771_end_0 = const()[name = string("op_1771_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1771_end_mask_0 = const()[name = string("op_1771_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1771_cast_fp16 = slice_by_index(begin = var_1771_begin_0, end = var_1771_end_0, end_mask = var_1771_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1771_cast_fp16")];
            tensor<int32, [4]> var_1786_begin_0 = const()[name = string("op_1786_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1786_end_0 = const()[name = string("op_1786_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1786_end_mask_0 = const()[name = string("op_1786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1786_cast_fp16 = slice_by_index(begin = var_1786_begin_0, end = var_1786_end_0, end_mask = var_1786_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1788 = const()[name = string("op_1788"), val = int32(1)];
            bool mixed_3_interleave_0 = const()[name = string("mixed_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_3_cast_fp16 = concat(axis = var_1788, interleave = mixed_3_interleave_0, values = (var_1771_cast_fp16, var_1786_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1792_perm_0 = const()[name = string("op_1792_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1792_cast_fp16 = transpose(perm = var_1792_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1796, x = var_1792_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315448064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317545280))))[name = string("model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_11_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1807_promoted_to_fp16 = const()[name = string("op_1807_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1809_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1807_promoted_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1811_axes_0 = const()[name = string("op_1811_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1811_keep_dims_0 = const()[name = string("op_1811_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1811_cast_fp16 = reduce_mean(axes = var_1811_axes_0, keep_dims = var_1811_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1811_cast_fp16")];
            fp16 var_1812_to_fp16 = const()[name = string("op_1812_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1813_cast_fp16 = add(x = var_1811_cast_fp16, y = var_1812_to_fp16)[name = string("op_1813_cast_fp16")];
            fp32 var_1814_epsilon_0 = const()[name = string("op_1814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1814_cast_fp16 = rsqrt(epsilon = var_1814_epsilon_0, x = var_1813_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1815_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1814_cast_fp16)[name = string("op_1815_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317547392)))];
            tensor<fp16, [1, 16, 1024]> input_5_cast_fp16 = mul(x = var_1815_cast_fp16, y = model_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317549504))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320695296))))[name = string("model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320701504)))];
            tensor<fp16, [1, 16, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_1820_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320707712))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323853504))))[name = string("model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_9_cast_fp16 = mul(x = var_1820_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323859712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327005504))))[name = string("model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_6_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1834_promoted_to_fp16 = const()[name = string("op_1834_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1836_cast_fp16 = pow(x = x_13_cast_fp16, y = var_1834_promoted_to_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [1]> var_1838_axes_0 = const()[name = string("op_1838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1838_keep_dims_0 = const()[name = string("op_1838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = reduce_mean(axes = var_1838_axes_0, keep_dims = var_1838_keep_dims_0, x = var_1836_cast_fp16)[name = string("op_1838_cast_fp16")];
            fp16 var_1839_to_fp16 = const()[name = string("op_1839_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1840_cast_fp16 = add(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            fp32 var_1841_epsilon_0 = const()[name = string("op_1841_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = rsqrt(epsilon = var_1841_epsilon_0, x = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1842_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327007616)))];
            tensor<fp16, [1, 16, 1024]> input_11_cast_fp16 = mul(x = var_1842_cast_fp16, y = model_layers_1_input_layernorm_weight_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327009728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329106944))))[name = string("model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_1851 = const()[name = string("op_1851"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1852_cast_fp16 = reshape(shape = var_1851, x = linear_7_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x_15_perm_0 = const()[name = string("x_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_15_cast_fp16 = transpose(perm = x_15_perm_0, x = var_1852_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1862_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [1]> var_1864_axes_0 = const()[name = string("op_1864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1864_keep_dims_0 = const()[name = string("op_1864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1864_cast_fp16 = reduce_mean(axes = var_1864_axes_0, keep_dims = var_1864_keep_dims_0, x = var_1862_cast_fp16)[name = string("op_1864_cast_fp16")];
            fp16 var_1865_to_fp16 = const()[name = string("op_1865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1866_cast_fp16 = add(x = var_1864_cast_fp16, y = var_1865_to_fp16)[name = string("op_1866_cast_fp16")];
            fp32 var_1867_epsilon_0 = const()[name = string("op_1867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1867_cast_fp16 = rsqrt(epsilon = var_1867_epsilon_0, x = var_1866_cast_fp16)[name = string("op_1867_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1868_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1867_cast_fp16)[name = string("op_1868_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111104)))];
            tensor<fp16, [1, 16, 16, 128]> x_19_cast_fp16 = mul(x = var_1868_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330160064))))[name = string("model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1877 = const()[name = string("op_1877"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1878_cast_fp16 = reshape(shape = var_1877, x = linear_8_cast_fp16)[name = string("op_1878_cast_fp16")];
            tensor<int32, [4]> x_17_perm_0 = const()[name = string("x_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1886_promoted_to_fp16 = const()[name = string("op_1886_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = var_1878_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1888_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1886_promoted_to_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [1]> var_1890_axes_0 = const()[name = string("op_1890_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1890_keep_dims_0 = const()[name = string("op_1890_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1890_cast_fp16 = reduce_mean(axes = var_1890_axes_0, keep_dims = var_1890_keep_dims_0, x = var_1888_cast_fp16)[name = string("op_1890_cast_fp16")];
            fp16 var_1891_to_fp16 = const()[name = string("op_1891_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1892_cast_fp16 = add(x = var_1890_cast_fp16, y = var_1891_to_fp16)[name = string("op_1892_cast_fp16")];
            fp32 var_1893_epsilon_0 = const()[name = string("op_1893_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1893_cast_fp16 = rsqrt(epsilon = var_1893_epsilon_0, x = var_1892_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1894_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1893_cast_fp16)[name = string("op_1894_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162176)))];
            tensor<fp16, [1, 8, 16, 128]> x_21_cast_fp16 = mul(x = var_1894_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331211136))))[name = string("model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1903 = const()[name = string("op_1903"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1904_cast_fp16 = reshape(shape = var_1903, x = linear_9_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_5_begin_0 = const()[name = string("first_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_5_end_0 = const()[name = string("first_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_5_end_mask_0 = const()[name = string("first_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_5_cast_fp16 = slice_by_index(begin = first_5_begin_0, end = first_5_end_0, end_mask = first_5_end_mask_0, x = x_19_cast_fp16)[name = string("first_5_cast_fp16")];
            tensor<int32, [4]> second_5_begin_0 = const()[name = string("second_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_5_end_0 = const()[name = string("second_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_5_end_mask_0 = const()[name = string("second_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_5_cast_fp16 = slice_by_index(begin = second_5_begin_0, end = second_5_end_0, end_mask = second_5_end_mask_0, x = x_19_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1918_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1919_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1921_cast_fp16 = sub(x = var_1918_cast_fp16, y = var_1919_cast_fp16)[name = string("op_1921_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1922_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1923_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1925_cast_fp16 = add(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1925_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(-1)];
            bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_3_cast_fp16 = concat(axis = var_1927, interleave = q_3_interleave_0, values = (var_1921_cast_fp16, var_1925_cast_fp16))[name = string("q_3_cast_fp16")];
            tensor<int32, [4]> first_7_begin_0 = const()[name = string("first_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_7_end_0 = const()[name = string("first_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_7_end_mask_0 = const()[name = string("first_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_7_cast_fp16 = slice_by_index(begin = first_7_begin_0, end = first_7_end_0, end_mask = first_7_end_mask_0, x = x_21_cast_fp16)[name = string("first_7_cast_fp16")];
            tensor<int32, [4]> second_7_begin_0 = const()[name = string("second_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_7_end_0 = const()[name = string("second_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_7_end_mask_0 = const()[name = string("second_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_7_cast_fp16 = slice_by_index(begin = second_7_begin_0, end = second_7_end_0, end_mask = second_7_end_mask_0, x = x_21_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1939_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1940_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1942_cast_fp16 = sub(x = var_1939_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1943_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1943_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1944_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1944_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1946_cast_fp16 = add(x = var_1943_cast_fp16, y = var_1944_cast_fp16)[name = string("op_1946_cast_fp16")];
            int32 var_1948 = const()[name = string("op_1948"), val = int32(-1)];
            bool k_3_interleave_0 = const()[name = string("k_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_3_cast_fp16 = concat(axis = var_1948, interleave = k_3_interleave_0, values = (var_1942_cast_fp16, var_1946_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_1959_begin_0 = const()[name = string("op_1959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1959_end_0 = const()[name = string("op_1959_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1959_end_mask_0 = const()[name = string("op_1959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1959_stride_0 = const()[name = string("op_1959_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, stride = var_1959_stride_0, x = q_3_cast_fp16)[name = string("op_1959_cast_fp16")];
            tensor<int32, [4]> var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1969_stride_0 = const()[name = string("op_1969_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, stride = var_1969_stride_0, x = q_3_cast_fp16)[name = string("op_1969_cast_fp16")];
            int32 var_1971 = const()[name = string("op_1971"), val = int32(2)];
            bool stacked_3_interleave_0 = const()[name = string("stacked_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_3_cast_fp16 = concat(axis = var_1971, interleave = stacked_3_interleave_0, values = (var_1959_cast_fp16, var_1969_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_1977_begin_0 = const()[name = string("op_1977_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1977_end_0 = const()[name = string("op_1977_end_0"), val = tensor<int32, [4]>([2, 8, 4112, 128])];
            tensor<bool, [4]> var_1977_end_mask_0 = const()[name = string("op_1977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, x = read_state_0)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> var_1987_begin_0 = const()[name = string("op_1987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1987_end_0 = const()[name = string("op_1987_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_1987_end_mask_0 = const()[name = string("op_1987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, x = var_1977_cast_fp16)[name = string("op_1987_cast_fp16")];
            bool var_1996_transpose_x_1 = const()[name = string("op_1996_transpose_x_1"), val = bool(false)];
            bool var_1996_transpose_y_1 = const()[name = string("op_1996_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_1996_cast_fp16 = matmul(transpose_x = var_1996_transpose_x_1, transpose_y = var_1996_transpose_y_1, x = var_1987_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_1996_cast_fp16")];
            tensor<int32, [4]> var_1999_perm_0 = const()[name = string("op_1999_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_1999_cast_fp16 = transpose(perm = var_1999_perm_0, x = var_1996_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 4096]> var_2001_cast_fp16 = mul(x = var_1999_cast_fp16, y = var_2000_to_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_3_cast_fp16 = add(x = var_2001_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2007_transpose_x_1 = const()[name = string("op_2007_transpose_x_1"), val = bool(false)];
            bool var_2007_transpose_y_1 = const()[name = string("op_2007_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2007_cast_fp16 = matmul(transpose_x = var_2007_transpose_x_1, transpose_y = var_2007_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2007_cast_fp16")];
            fp16 var_2008_to_fp16 = const()[name = string("op_2008_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2009_cast_fp16 = mul(x = var_2007_cast_fp16, y = var_2008_to_fp16)[name = string("op_2009_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2009_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2013 = const()[name = string("op_2013"), val = int32(-1)];
            bool var_2014_interleave_0 = const()[name = string("op_2014_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2014_cast_fp16 = concat(axis = var_2013, interleave = var_2014_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2014_cast_fp16")];
            int32 var_2015 = const()[name = string("op_2015"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_3_cast_fp16 = softmax(axis = var_2015, x = var_2014_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2022_begin_0 = const()[name = string("op_2022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2022_end_0 = const()[name = string("op_2022_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2022_end_mask_0 = const()[name = string("op_2022_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2022_cast_fp16 = slice_by_index(begin = var_2022_begin_0, end = var_2022_end_0, end_mask = var_2022_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<int32, [4]> var_2027_begin_0 = const()[name = string("op_2027_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2027_end_0 = const()[name = string("op_2027_end_0"), val = tensor<int32, [4]>([2, 8, 4112, 128])];
            tensor<bool, [4]> var_2027_end_mask_0 = const()[name = string("op_2027_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2027_cast_fp16 = slice_by_index(begin = var_2027_begin_0, end = var_2027_end_0, end_mask = var_2027_end_mask_0, x = read_state_1)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2037_begin_0 = const()[name = string("op_2037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2037_end_0 = const()[name = string("op_2037_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2037_end_mask_0 = const()[name = string("op_2037_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2037_cast_fp16 = slice_by_index(begin = var_2037_begin_0, end = var_2037_end_0, end_mask = var_2037_end_mask_0, x = var_2027_cast_fp16)[name = string("op_2037_cast_fp16")];
            bool var_2043_transpose_x_0 = const()[name = string("op_2043_transpose_x_0"), val = bool(false)];
            bool var_2043_transpose_y_0 = const()[name = string("op_2043_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2043_cast_fp16 = matmul(transpose_x = var_2043_transpose_x_0, transpose_y = var_2043_transpose_y_0, x = var_2022_cast_fp16, y = var_2037_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [4]> var_2048_begin_0 = const()[name = string("op_2048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2048_end_0 = const()[name = string("op_2048_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2048_end_mask_0 = const()[name = string("op_2048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2048_cast_fp16")];
            bool var_2049_transpose_x_0 = const()[name = string("op_2049_transpose_x_0"), val = bool(false)];
            bool var_2049_transpose_y_0 = const()[name = string("op_2049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1904_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2049_cast_fp16 = matmul(transpose_x = var_2049_transpose_x_0, transpose_y = var_2049_transpose_y_0, x = var_2048_cast_fp16, y = v_3_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2049_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2066_begin_0 = const()[name = string("op_2066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2066_end_0 = const()[name = string("op_2066_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2066_end_mask_0 = const()[name = string("op_2066_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2066_cast_fp16 = slice_by_index(begin = var_2066_begin_0, end = var_2066_end_0, end_mask = var_2066_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<int32, [4]> var_2081_begin_0 = const()[name = string("op_2081_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2081_end_0 = const()[name = string("op_2081_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2081_end_mask_0 = const()[name = string("op_2081_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2081_cast_fp16 = slice_by_index(begin = var_2081_begin_0, end = var_2081_end_0, end_mask = var_2081_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2081_cast_fp16")];
            int32 var_2083 = const()[name = string("op_2083"), val = int32(1)];
            bool mixed_7_interleave_0 = const()[name = string("mixed_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_7_cast_fp16 = concat(axis = var_2083, interleave = mixed_7_interleave_0, values = (var_2066_cast_fp16, var_2081_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2087_perm_0 = const()[name = string("op_2087_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2091 = const()[name = string("op_2091"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2087_cast_fp16 = transpose(perm = var_2087_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2091, x = var_2087_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331213248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333310464))))[name = string("model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = add(x = x_13_cast_fp16, y = linear_10_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_2102_promoted_to_fp16 = const()[name = string("op_2102_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2104_cast_fp16 = pow(x = x_23_cast_fp16, y = var_2102_promoted_to_fp16)[name = string("op_2104_cast_fp16")];
            tensor<int32, [1]> var_2106_axes_0 = const()[name = string("op_2106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2106_keep_dims_0 = const()[name = string("op_2106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2106_cast_fp16 = reduce_mean(axes = var_2106_axes_0, keep_dims = var_2106_keep_dims_0, x = var_2104_cast_fp16)[name = string("op_2106_cast_fp16")];
            fp16 var_2107_to_fp16 = const()[name = string("op_2107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2108_cast_fp16 = add(x = var_2106_cast_fp16, y = var_2107_to_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2110_cast_fp16 = mul(x = x_23_cast_fp16, y = var_2109_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333312576)))];
            tensor<fp16, [1, 16, 1024]> input_15_cast_fp16 = mul(x = var_2110_cast_fp16, y = model_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333314688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336460480))))[name = string("model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2115_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336466688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339612480))))[name = string("model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_19_cast_fp16 = mul(x = var_2115_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339618688))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342764480))))[name = string("model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_13_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_2129_promoted_to_fp16 = const()[name = string("op_2129_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2131_cast_fp16 = pow(x = x_25_cast_fp16, y = var_2129_promoted_to_fp16)[name = string("op_2131_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2133_keep_dims_0 = const()[name = string("op_2133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2133_cast_fp16 = reduce_mean(axes = var_2133_axes_0, keep_dims = var_2133_keep_dims_0, x = var_2131_cast_fp16)[name = string("op_2133_cast_fp16")];
            fp16 var_2134_to_fp16 = const()[name = string("op_2134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2135_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = string("op_2135_cast_fp16")];
            fp32 var_2136_epsilon_0 = const()[name = string("op_2136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2136_cast_fp16 = rsqrt(epsilon = var_2136_epsilon_0, x = var_2135_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2137_cast_fp16 = mul(x = x_25_cast_fp16, y = var_2136_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342766592)))];
            tensor<fp16, [1, 16, 1024]> input_21_cast_fp16 = mul(x = var_2137_cast_fp16, y = model_layers_2_input_layernorm_weight_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342768704))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344865920))))[name = string("model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_2146 = const()[name = string("op_2146"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2147_cast_fp16 = reshape(shape = var_2146, x = linear_14_cast_fp16)[name = string("op_2147_cast_fp16")];
            tensor<int32, [4]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2155_promoted_to_fp16 = const()[name = string("op_2155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = var_2147_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2157_cast_fp16 = pow(x = x_27_cast_fp16, y = var_2155_promoted_to_fp16)[name = string("op_2157_cast_fp16")];
            tensor<int32, [1]> var_2159_axes_0 = const()[name = string("op_2159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2159_keep_dims_0 = const()[name = string("op_2159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2159_cast_fp16 = reduce_mean(axes = var_2159_axes_0, keep_dims = var_2159_keep_dims_0, x = var_2157_cast_fp16)[name = string("op_2159_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2161_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
            fp32 var_2162_epsilon_0 = const()[name = string("op_2162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2162_cast_fp16 = rsqrt(epsilon = var_2162_epsilon_0, x = var_2161_cast_fp16)[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2163_cast_fp16 = mul(x = x_27_cast_fp16, y = var_2162_cast_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870080)))];
            tensor<fp16, [1, 16, 16, 128]> x_31_cast_fp16 = mul(x = var_2163_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345919040))))[name = string("model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2172 = const()[name = string("op_2172"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2173_cast_fp16 = reshape(shape = var_2172, x = linear_15_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2181_promoted_to_fp16 = const()[name = string("op_2181_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_2173_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2183_cast_fp16 = pow(x = x_29_cast_fp16, y = var_2181_promoted_to_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [1]> var_2185_axes_0 = const()[name = string("op_2185_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2185_keep_dims_0 = const()[name = string("op_2185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2185_cast_fp16 = reduce_mean(axes = var_2185_axes_0, keep_dims = var_2185_keep_dims_0, x = var_2183_cast_fp16)[name = string("op_2185_cast_fp16")];
            fp16 var_2186_to_fp16 = const()[name = string("op_2186_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2187_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2186_to_fp16)[name = string("op_2187_cast_fp16")];
            fp32 var_2188_epsilon_0 = const()[name = string("op_2188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2188_cast_fp16 = rsqrt(epsilon = var_2188_epsilon_0, x = var_2187_cast_fp16)[name = string("op_2188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2189_cast_fp16 = mul(x = x_29_cast_fp16, y = var_2188_cast_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921152)))];
            tensor<fp16, [1, 8, 16, 128]> x_33_cast_fp16 = mul(x = var_2189_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346970112))))[name = string("model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2198 = const()[name = string("op_2198"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2199_cast_fp16 = reshape(shape = var_2198, x = linear_16_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_9_begin_0 = const()[name = string("first_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_9_end_0 = const()[name = string("first_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_9_end_mask_0 = const()[name = string("first_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_9_cast_fp16 = slice_by_index(begin = first_9_begin_0, end = first_9_end_0, end_mask = first_9_end_mask_0, x = x_31_cast_fp16)[name = string("first_9_cast_fp16")];
            tensor<int32, [4]> second_9_begin_0 = const()[name = string("second_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_9_end_0 = const()[name = string("second_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_9_end_mask_0 = const()[name = string("second_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_9_cast_fp16 = slice_by_index(begin = second_9_begin_0, end = second_9_end_0, end_mask = second_9_end_mask_0, x = x_31_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2213_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2213_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2214_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2216_cast_fp16 = sub(x = var_2213_cast_fp16, y = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2217_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2218_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2220_cast_fp16 = add(x = var_2217_cast_fp16, y = var_2218_cast_fp16)[name = string("op_2220_cast_fp16")];
            int32 var_2222 = const()[name = string("op_2222"), val = int32(-1)];
            bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_5_cast_fp16 = concat(axis = var_2222, interleave = q_5_interleave_0, values = (var_2216_cast_fp16, var_2220_cast_fp16))[name = string("q_5_cast_fp16")];
            tensor<int32, [4]> first_11_begin_0 = const()[name = string("first_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_11_end_0 = const()[name = string("first_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_11_end_mask_0 = const()[name = string("first_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_11_cast_fp16 = slice_by_index(begin = first_11_begin_0, end = first_11_end_0, end_mask = first_11_end_mask_0, x = x_33_cast_fp16)[name = string("first_11_cast_fp16")];
            tensor<int32, [4]> second_11_begin_0 = const()[name = string("second_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_11_end_0 = const()[name = string("second_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_11_end_mask_0 = const()[name = string("second_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_11_cast_fp16 = slice_by_index(begin = second_11_begin_0, end = second_11_end_0, end_mask = second_11_end_mask_0, x = x_33_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2234_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2235_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2237_cast_fp16 = sub(x = var_2234_cast_fp16, y = var_2235_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2238_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2238_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2239_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2241_cast_fp16 = add(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = string("op_2241_cast_fp16")];
            int32 var_2243 = const()[name = string("op_2243"), val = int32(-1)];
            bool k_5_interleave_0 = const()[name = string("k_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_5_cast_fp16 = concat(axis = var_2243, interleave = k_5_interleave_0, values = (var_2237_cast_fp16, var_2241_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2254_begin_0 = const()[name = string("op_2254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2254_end_0 = const()[name = string("op_2254_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2254_end_mask_0 = const()[name = string("op_2254_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2254_stride_0 = const()[name = string("op_2254_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2254_cast_fp16 = slice_by_index(begin = var_2254_begin_0, end = var_2254_end_0, end_mask = var_2254_end_mask_0, stride = var_2254_stride_0, x = q_5_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<int32, [4]> var_2264_begin_0 = const()[name = string("op_2264_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2264_end_0 = const()[name = string("op_2264_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2264_end_mask_0 = const()[name = string("op_2264_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2264_stride_0 = const()[name = string("op_2264_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2264_cast_fp16 = slice_by_index(begin = var_2264_begin_0, end = var_2264_end_0, end_mask = var_2264_end_mask_0, stride = var_2264_stride_0, x = q_5_cast_fp16)[name = string("op_2264_cast_fp16")];
            int32 var_2266 = const()[name = string("op_2266"), val = int32(2)];
            bool stacked_5_interleave_0 = const()[name = string("stacked_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_5_cast_fp16 = concat(axis = var_2266, interleave = stacked_5_interleave_0, values = (var_2254_cast_fp16, var_2264_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2272_begin_0 = const()[name = string("op_2272_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2272_end_0 = const()[name = string("op_2272_end_0"), val = tensor<int32, [4]>([3, 8, 4112, 128])];
            tensor<bool, [4]> var_2272_end_mask_0 = const()[name = string("op_2272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2272_cast_fp16 = slice_by_index(begin = var_2272_begin_0, end = var_2272_end_0, end_mask = var_2272_end_mask_0, x = read_state_0)[name = string("op_2272_cast_fp16")];
            tensor<int32, [4]> var_2282_begin_0 = const()[name = string("op_2282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2282_end_0 = const()[name = string("op_2282_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2282_end_mask_0 = const()[name = string("op_2282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2282_cast_fp16 = slice_by_index(begin = var_2282_begin_0, end = var_2282_end_0, end_mask = var_2282_end_mask_0, x = var_2272_cast_fp16)[name = string("op_2282_cast_fp16")];
            bool var_2291_transpose_x_1 = const()[name = string("op_2291_transpose_x_1"), val = bool(false)];
            bool var_2291_transpose_y_1 = const()[name = string("op_2291_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2291_cast_fp16 = matmul(transpose_x = var_2291_transpose_x_1, transpose_y = var_2291_transpose_y_1, x = var_2282_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2294_perm_0 = const()[name = string("op_2294_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2295_to_fp16 = const()[name = string("op_2295_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2294_cast_fp16 = transpose(perm = var_2294_perm_0, x = var_2291_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 4096]> var_2296_cast_fp16 = mul(x = var_2294_cast_fp16, y = var_2295_to_fp16)[name = string("op_2296_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_5_cast_fp16 = add(x = var_2296_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2302_transpose_x_1 = const()[name = string("op_2302_transpose_x_1"), val = bool(false)];
            bool var_2302_transpose_y_1 = const()[name = string("op_2302_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2302_cast_fp16 = matmul(transpose_x = var_2302_transpose_x_1, transpose_y = var_2302_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp16 var_2303_to_fp16 = const()[name = string("op_2303_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2304_cast_fp16 = mul(x = var_2302_cast_fp16, y = var_2303_to_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2304_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2308 = const()[name = string("op_2308"), val = int32(-1)];
            bool var_2309_interleave_0 = const()[name = string("op_2309_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2309_cast_fp16 = concat(axis = var_2308, interleave = var_2309_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2309_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_5_cast_fp16 = softmax(axis = var_2310, x = var_2309_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<int32, [4]> var_2322_begin_0 = const()[name = string("op_2322_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2322_end_0 = const()[name = string("op_2322_end_0"), val = tensor<int32, [4]>([3, 8, 4112, 128])];
            tensor<bool, [4]> var_2322_end_mask_0 = const()[name = string("op_2322_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = read_state_1)[name = string("op_2322_cast_fp16")];
            tensor<int32, [4]> var_2332_begin_0 = const()[name = string("op_2332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2332_end_0 = const()[name = string("op_2332_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2332_end_mask_0 = const()[name = string("op_2332_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2332_cast_fp16 = slice_by_index(begin = var_2332_begin_0, end = var_2332_end_0, end_mask = var_2332_end_mask_0, x = var_2322_cast_fp16)[name = string("op_2332_cast_fp16")];
            bool var_2338_transpose_x_0 = const()[name = string("op_2338_transpose_x_0"), val = bool(false)];
            bool var_2338_transpose_y_0 = const()[name = string("op_2338_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2338_cast_fp16 = matmul(transpose_x = var_2338_transpose_x_0, transpose_y = var_2338_transpose_y_0, x = var_2317_cast_fp16, y = var_2332_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<int32, [4]> var_2343_begin_0 = const()[name = string("op_2343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2343_end_0 = const()[name = string("op_2343_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2343_end_mask_0 = const()[name = string("op_2343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2343_cast_fp16 = slice_by_index(begin = var_2343_begin_0, end = var_2343_end_0, end_mask = var_2343_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2343_cast_fp16")];
            bool var_2344_transpose_x_0 = const()[name = string("op_2344_transpose_x_0"), val = bool(false)];
            bool var_2344_transpose_y_0 = const()[name = string("op_2344_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2199_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2344_cast_fp16 = matmul(transpose_x = var_2344_transpose_x_0, transpose_y = var_2344_transpose_y_0, x = var_2343_cast_fp16, y = v_5_cast_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2338_cast_fp16, y = var_2344_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2361_begin_0 = const()[name = string("op_2361_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2361_end_0 = const()[name = string("op_2361_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2361_end_mask_0 = const()[name = string("op_2361_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2361_cast_fp16")];
            tensor<int32, [4]> var_2376_begin_0 = const()[name = string("op_2376_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2376_end_0 = const()[name = string("op_2376_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2376_end_mask_0 = const()[name = string("op_2376_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2376_cast_fp16 = slice_by_index(begin = var_2376_begin_0, end = var_2376_end_0, end_mask = var_2376_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(1)];
            bool mixed_11_interleave_0 = const()[name = string("mixed_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_11_cast_fp16 = concat(axis = var_2378, interleave = mixed_11_interleave_0, values = (var_2361_cast_fp16, var_2376_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2382_perm_0 = const()[name = string("op_2382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2382_cast_fp16 = transpose(perm = var_2382_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346972224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349069440))))[name = string("model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_35_cast_fp16 = add(x = x_25_cast_fp16, y = linear_17_cast_fp16)[name = string("x_35_cast_fp16")];
            fp16 var_2397_promoted_to_fp16 = const()[name = string("op_2397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2399_cast_fp16 = pow(x = x_35_cast_fp16, y = var_2397_promoted_to_fp16)[name = string("op_2399_cast_fp16")];
            tensor<int32, [1]> var_2401_axes_0 = const()[name = string("op_2401_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2401_keep_dims_0 = const()[name = string("op_2401_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2401_cast_fp16 = reduce_mean(axes = var_2401_axes_0, keep_dims = var_2401_keep_dims_0, x = var_2399_cast_fp16)[name = string("op_2401_cast_fp16")];
            fp16 var_2402_to_fp16 = const()[name = string("op_2402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2403_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2402_to_fp16)[name = string("op_2403_cast_fp16")];
            fp32 var_2404_epsilon_0 = const()[name = string("op_2404_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2404_cast_fp16 = rsqrt(epsilon = var_2404_epsilon_0, x = var_2403_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2405_cast_fp16 = mul(x = x_35_cast_fp16, y = var_2404_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349071552)))];
            tensor<fp16, [1, 16, 1024]> input_25_cast_fp16 = mul(x = var_2405_cast_fp16, y = model_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349073664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352219456))))[name = string("model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2410_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352225664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355371456))))[name = string("model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_29_cast_fp16 = mul(x = var_2410_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355377664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358523456))))[name = string("model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_20_cast_fp16)[name = string("x_37_cast_fp16")];
            fp16 var_2424_promoted_to_fp16 = const()[name = string("op_2424_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2426_cast_fp16 = pow(x = x_37_cast_fp16, y = var_2424_promoted_to_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [1]> var_2428_axes_0 = const()[name = string("op_2428_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2428_keep_dims_0 = const()[name = string("op_2428_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2428_cast_fp16 = reduce_mean(axes = var_2428_axes_0, keep_dims = var_2428_keep_dims_0, x = var_2426_cast_fp16)[name = string("op_2428_cast_fp16")];
            fp16 var_2429_to_fp16 = const()[name = string("op_2429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2430_cast_fp16 = add(x = var_2428_cast_fp16, y = var_2429_to_fp16)[name = string("op_2430_cast_fp16")];
            fp32 var_2431_epsilon_0 = const()[name = string("op_2431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2431_cast_fp16 = rsqrt(epsilon = var_2431_epsilon_0, x = var_2430_cast_fp16)[name = string("op_2431_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2432_cast_fp16 = mul(x = x_37_cast_fp16, y = var_2431_cast_fp16)[name = string("op_2432_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358525568)))];
            tensor<fp16, [1, 16, 1024]> input_31_cast_fp16 = mul(x = var_2432_cast_fp16, y = model_layers_3_input_layernorm_weight_to_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358527680))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360624896))))[name = string("model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2442_cast_fp16 = reshape(shape = var_2441, x = linear_21_cast_fp16)[name = string("op_2442_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2450_promoted_to_fp16 = const()[name = string("op_2450_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2442_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2452_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2450_promoted_to_fp16)[name = string("op_2452_cast_fp16")];
            tensor<int32, [1]> var_2454_axes_0 = const()[name = string("op_2454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2454_keep_dims_0 = const()[name = string("op_2454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2454_cast_fp16 = reduce_mean(axes = var_2454_axes_0, keep_dims = var_2454_keep_dims_0, x = var_2452_cast_fp16)[name = string("op_2454_cast_fp16")];
            fp16 var_2455_to_fp16 = const()[name = string("op_2455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2456_cast_fp16 = add(x = var_2454_cast_fp16, y = var_2455_to_fp16)[name = string("op_2456_cast_fp16")];
            fp32 var_2457_epsilon_0 = const()[name = string("op_2457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2457_cast_fp16 = rsqrt(epsilon = var_2457_epsilon_0, x = var_2456_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2458_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629056)))];
            tensor<fp16, [1, 16, 16, 128]> x_43_cast_fp16 = mul(x = var_2458_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361678016))))[name = string("model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2468_cast_fp16 = reshape(shape = var_2467, x = linear_22_cast_fp16)[name = string("op_2468_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2476_promoted_to_fp16 = const()[name = string("op_2476_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2468_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2478_cast_fp16 = pow(x = x_41_cast_fp16, y = var_2476_promoted_to_fp16)[name = string("op_2478_cast_fp16")];
            tensor<int32, [1]> var_2480_axes_0 = const()[name = string("op_2480_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2480_keep_dims_0 = const()[name = string("op_2480_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2480_cast_fp16 = reduce_mean(axes = var_2480_axes_0, keep_dims = var_2480_keep_dims_0, x = var_2478_cast_fp16)[name = string("op_2480_cast_fp16")];
            fp16 var_2481_to_fp16 = const()[name = string("op_2481_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2482_cast_fp16 = add(x = var_2480_cast_fp16, y = var_2481_to_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2484_cast_fp16 = mul(x = x_41_cast_fp16, y = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680128)))];
            tensor<fp16, [1, 8, 16, 128]> x_45_cast_fp16 = mul(x = var_2484_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362729088))))[name = string("model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2493 = const()[name = string("op_2493"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2494_cast_fp16 = reshape(shape = var_2493, x = linear_23_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_13_begin_0 = const()[name = string("first_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_13_end_0 = const()[name = string("first_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_13_end_mask_0 = const()[name = string("first_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_13_cast_fp16 = slice_by_index(begin = first_13_begin_0, end = first_13_end_0, end_mask = first_13_end_mask_0, x = x_43_cast_fp16)[name = string("first_13_cast_fp16")];
            tensor<int32, [4]> second_13_begin_0 = const()[name = string("second_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_13_end_0 = const()[name = string("second_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_13_end_mask_0 = const()[name = string("second_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_13_cast_fp16 = slice_by_index(begin = second_13_begin_0, end = second_13_end_0, end_mask = second_13_end_mask_0, x = x_43_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2508_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2508_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2509_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2509_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2511_cast_fp16 = sub(x = var_2508_cast_fp16, y = var_2509_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2512_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2512_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2513_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2515_cast_fp16 = add(x = var_2512_cast_fp16, y = var_2513_cast_fp16)[name = string("op_2515_cast_fp16")];
            int32 var_2517 = const()[name = string("op_2517"), val = int32(-1)];
            bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_7_cast_fp16 = concat(axis = var_2517, interleave = q_7_interleave_0, values = (var_2511_cast_fp16, var_2515_cast_fp16))[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> first_15_begin_0 = const()[name = string("first_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_15_end_0 = const()[name = string("first_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_15_end_mask_0 = const()[name = string("first_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_15_cast_fp16 = slice_by_index(begin = first_15_begin_0, end = first_15_end_0, end_mask = first_15_end_mask_0, x = x_45_cast_fp16)[name = string("first_15_cast_fp16")];
            tensor<int32, [4]> second_15_begin_0 = const()[name = string("second_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_15_end_0 = const()[name = string("second_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_15_end_mask_0 = const()[name = string("second_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_15_cast_fp16 = slice_by_index(begin = second_15_begin_0, end = second_15_end_0, end_mask = second_15_end_mask_0, x = x_45_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2529_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2530_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2532_cast_fp16 = sub(x = var_2529_cast_fp16, y = var_2530_cast_fp16)[name = string("op_2532_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2533_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2534_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2534_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2536_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2536_cast_fp16")];
            int32 var_2538 = const()[name = string("op_2538"), val = int32(-1)];
            bool k_7_interleave_0 = const()[name = string("k_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_7_cast_fp16 = concat(axis = var_2538, interleave = k_7_interleave_0, values = (var_2532_cast_fp16, var_2536_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2549_begin_0 = const()[name = string("op_2549_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2549_end_0 = const()[name = string("op_2549_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2549_end_mask_0 = const()[name = string("op_2549_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2549_stride_0 = const()[name = string("op_2549_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2549_cast_fp16 = slice_by_index(begin = var_2549_begin_0, end = var_2549_end_0, end_mask = var_2549_end_mask_0, stride = var_2549_stride_0, x = q_7_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [4]> var_2559_begin_0 = const()[name = string("op_2559_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2559_end_0 = const()[name = string("op_2559_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2559_end_mask_0 = const()[name = string("op_2559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2559_stride_0 = const()[name = string("op_2559_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2559_cast_fp16 = slice_by_index(begin = var_2559_begin_0, end = var_2559_end_0, end_mask = var_2559_end_mask_0, stride = var_2559_stride_0, x = q_7_cast_fp16)[name = string("op_2559_cast_fp16")];
            int32 var_2561 = const()[name = string("op_2561"), val = int32(2)];
            bool stacked_7_interleave_0 = const()[name = string("stacked_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_7_cast_fp16 = concat(axis = var_2561, interleave = stacked_7_interleave_0, values = (var_2549_cast_fp16, var_2559_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2567_begin_0 = const()[name = string("op_2567_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2567_end_0 = const()[name = string("op_2567_end_0"), val = tensor<int32, [4]>([4, 8, 4112, 128])];
            tensor<bool, [4]> var_2567_end_mask_0 = const()[name = string("op_2567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2567_cast_fp16 = slice_by_index(begin = var_2567_begin_0, end = var_2567_end_0, end_mask = var_2567_end_mask_0, x = read_state_0)[name = string("op_2567_cast_fp16")];
            tensor<int32, [4]> var_2577_begin_0 = const()[name = string("op_2577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2577_end_0 = const()[name = string("op_2577_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2577_end_mask_0 = const()[name = string("op_2577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2577_cast_fp16 = slice_by_index(begin = var_2577_begin_0, end = var_2577_end_0, end_mask = var_2577_end_mask_0, x = var_2567_cast_fp16)[name = string("op_2577_cast_fp16")];
            bool var_2586_transpose_x_1 = const()[name = string("op_2586_transpose_x_1"), val = bool(false)];
            bool var_2586_transpose_y_1 = const()[name = string("op_2586_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2586_cast_fp16 = matmul(transpose_x = var_2586_transpose_x_1, transpose_y = var_2586_transpose_y_1, x = var_2577_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2586_cast_fp16")];
            tensor<int32, [4]> var_2589_perm_0 = const()[name = string("op_2589_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2590_to_fp16 = const()[name = string("op_2590_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2589_cast_fp16 = transpose(perm = var_2589_perm_0, x = var_2586_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 4096]> var_2591_cast_fp16 = mul(x = var_2589_cast_fp16, y = var_2590_to_fp16)[name = string("op_2591_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_7_cast_fp16 = add(x = var_2591_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2597_transpose_x_1 = const()[name = string("op_2597_transpose_x_1"), val = bool(false)];
            bool var_2597_transpose_y_1 = const()[name = string("op_2597_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2597_cast_fp16 = matmul(transpose_x = var_2597_transpose_x_1, transpose_y = var_2597_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2597_cast_fp16")];
            fp16 var_2598_to_fp16 = const()[name = string("op_2598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2599_cast_fp16 = mul(x = var_2597_cast_fp16, y = var_2598_to_fp16)[name = string("op_2599_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2599_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(-1)];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2604_cast_fp16 = concat(axis = var_2603, interleave = var_2604_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2604_cast_fp16")];
            int32 var_2605 = const()[name = string("op_2605"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_7_cast_fp16 = softmax(axis = var_2605, x = var_2604_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2612_begin_0 = const()[name = string("op_2612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2612_end_0 = const()[name = string("op_2612_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2612_end_mask_0 = const()[name = string("op_2612_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2612_cast_fp16 = slice_by_index(begin = var_2612_begin_0, end = var_2612_end_0, end_mask = var_2612_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2612_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([4, 8, 4112, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = read_state_1)[name = string("op_2617_cast_fp16")];
            tensor<int32, [4]> var_2627_begin_0 = const()[name = string("op_2627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2627_end_0 = const()[name = string("op_2627_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2627_end_mask_0 = const()[name = string("op_2627_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2627_cast_fp16 = slice_by_index(begin = var_2627_begin_0, end = var_2627_end_0, end_mask = var_2627_end_mask_0, x = var_2617_cast_fp16)[name = string("op_2627_cast_fp16")];
            bool var_2633_transpose_x_0 = const()[name = string("op_2633_transpose_x_0"), val = bool(false)];
            bool var_2633_transpose_y_0 = const()[name = string("op_2633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2633_cast_fp16 = matmul(transpose_x = var_2633_transpose_x_0, transpose_y = var_2633_transpose_y_0, x = var_2612_cast_fp16, y = var_2627_cast_fp16)[name = string("op_2633_cast_fp16")];
            tensor<int32, [4]> var_2638_begin_0 = const()[name = string("op_2638_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2638_end_0 = const()[name = string("op_2638_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2638_end_mask_0 = const()[name = string("op_2638_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2638_cast_fp16 = slice_by_index(begin = var_2638_begin_0, end = var_2638_end_0, end_mask = var_2638_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2638_cast_fp16")];
            bool var_2639_transpose_x_0 = const()[name = string("op_2639_transpose_x_0"), val = bool(false)];
            bool var_2639_transpose_y_0 = const()[name = string("op_2639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2494_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2639_cast_fp16 = matmul(transpose_x = var_2639_transpose_x_0, transpose_y = var_2639_transpose_y_0, x = var_2638_cast_fp16, y = v_7_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2633_cast_fp16, y = var_2639_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2656_cast_fp16")];
            tensor<int32, [4]> var_2671_begin_0 = const()[name = string("op_2671_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2671_end_0 = const()[name = string("op_2671_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2671_end_mask_0 = const()[name = string("op_2671_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2671_cast_fp16 = slice_by_index(begin = var_2671_begin_0, end = var_2671_end_0, end_mask = var_2671_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2671_cast_fp16")];
            int32 var_2673 = const()[name = string("op_2673"), val = int32(1)];
            bool mixed_15_interleave_0 = const()[name = string("mixed_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_15_cast_fp16 = concat(axis = var_2673, interleave = mixed_15_interleave_0, values = (var_2656_cast_fp16, var_2671_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2681 = const()[name = string("op_2681"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2681, x = var_2677_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362731200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364828416))))[name = string("model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = linear_24_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2692_promoted_to_fp16 = const()[name = string("op_2692_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2694_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2692_promoted_to_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [1]> var_2696_axes_0 = const()[name = string("op_2696_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2696_keep_dims_0 = const()[name = string("op_2696_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2696_cast_fp16 = reduce_mean(axes = var_2696_axes_0, keep_dims = var_2696_keep_dims_0, x = var_2694_cast_fp16)[name = string("op_2696_cast_fp16")];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2698_cast_fp16 = add(x = var_2696_cast_fp16, y = var_2697_to_fp16)[name = string("op_2698_cast_fp16")];
            fp32 var_2699_epsilon_0 = const()[name = string("op_2699_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2699_cast_fp16 = rsqrt(epsilon = var_2699_epsilon_0, x = var_2698_cast_fp16)[name = string("op_2699_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2700_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2699_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364830528)))];
            tensor<fp16, [1, 16, 1024]> input_35_cast_fp16 = mul(x = var_2700_cast_fp16, y = model_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364832640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367978432))))[name = string("model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2705_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367984640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371130432))))[name = string("model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_39_cast_fp16 = mul(x = var_2705_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371136640))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374282432))))[name = string("model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_27_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_2719_promoted_to_fp16 = const()[name = string("op_2719_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2721_cast_fp16 = pow(x = x_49_cast_fp16, y = var_2719_promoted_to_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]>([-1])];
            bool var_2723_keep_dims_0 = const()[name = string("op_2723_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2723_cast_fp16 = reduce_mean(axes = var_2723_axes_0, keep_dims = var_2723_keep_dims_0, x = var_2721_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 var_2724_to_fp16 = const()[name = string("op_2724_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2725_cast_fp16 = add(x = var_2723_cast_fp16, y = var_2724_to_fp16)[name = string("op_2725_cast_fp16")];
            fp32 var_2726_epsilon_0 = const()[name = string("op_2726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2726_cast_fp16 = rsqrt(epsilon = var_2726_epsilon_0, x = var_2725_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2727_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2726_cast_fp16)[name = string("op_2727_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374284544)))];
            tensor<fp16, [1, 16, 1024]> input_41_cast_fp16 = mul(x = var_2727_cast_fp16, y = model_layers_4_input_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374286656))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376383872))))[name = string("model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_2736 = const()[name = string("op_2736"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2737_cast_fp16 = reshape(shape = var_2736, x = linear_28_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<int32, [4]> x_51_perm_0 = const()[name = string("x_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2745_promoted_to_fp16 = const()[name = string("op_2745_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_51_cast_fp16 = transpose(perm = x_51_perm_0, x = var_2737_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2747_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2745_promoted_to_fp16)[name = string("op_2747_cast_fp16")];
            tensor<int32, [1]> var_2749_axes_0 = const()[name = string("op_2749_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2749_keep_dims_0 = const()[name = string("op_2749_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2749_cast_fp16 = reduce_mean(axes = var_2749_axes_0, keep_dims = var_2749_keep_dims_0, x = var_2747_cast_fp16)[name = string("op_2749_cast_fp16")];
            fp16 var_2750_to_fp16 = const()[name = string("op_2750_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2751_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2750_to_fp16)[name = string("op_2751_cast_fp16")];
            fp32 var_2752_epsilon_0 = const()[name = string("op_2752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2752_cast_fp16 = rsqrt(epsilon = var_2752_epsilon_0, x = var_2751_cast_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2753_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2752_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388032)))];
            tensor<fp16, [1, 16, 16, 128]> x_55_cast_fp16 = mul(x = var_2753_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377436992))))[name = string("model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2762 = const()[name = string("op_2762"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2763_cast_fp16 = reshape(shape = var_2762, x = linear_29_cast_fp16)[name = string("op_2763_cast_fp16")];
            tensor<int32, [4]> x_53_perm_0 = const()[name = string("x_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2771_promoted_to_fp16 = const()[name = string("op_2771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_53_cast_fp16 = transpose(perm = x_53_perm_0, x = var_2763_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2773_cast_fp16 = pow(x = x_53_cast_fp16, y = var_2771_promoted_to_fp16)[name = string("op_2773_cast_fp16")];
            tensor<int32, [1]> var_2775_axes_0 = const()[name = string("op_2775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2775_keep_dims_0 = const()[name = string("op_2775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2775_cast_fp16 = reduce_mean(axes = var_2775_axes_0, keep_dims = var_2775_keep_dims_0, x = var_2773_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp16 var_2776_to_fp16 = const()[name = string("op_2776_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2777_cast_fp16 = add(x = var_2775_cast_fp16, y = var_2776_to_fp16)[name = string("op_2777_cast_fp16")];
            fp32 var_2778_epsilon_0 = const()[name = string("op_2778_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2778_cast_fp16 = rsqrt(epsilon = var_2778_epsilon_0, x = var_2777_cast_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2779_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2779_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439104)))];
            tensor<fp16, [1, 8, 16, 128]> x_57_cast_fp16 = mul(x = var_2779_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488064))))[name = string("model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2788 = const()[name = string("op_2788"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2789_cast_fp16 = reshape(shape = var_2788, x = linear_30_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_17_begin_0 = const()[name = string("first_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_17_end_0 = const()[name = string("first_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_17_end_mask_0 = const()[name = string("first_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_17_cast_fp16 = slice_by_index(begin = first_17_begin_0, end = first_17_end_0, end_mask = first_17_end_mask_0, x = x_55_cast_fp16)[name = string("first_17_cast_fp16")];
            tensor<int32, [4]> second_17_begin_0 = const()[name = string("second_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_17_end_0 = const()[name = string("second_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_17_end_mask_0 = const()[name = string("second_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_17_cast_fp16 = slice_by_index(begin = second_17_begin_0, end = second_17_end_0, end_mask = second_17_end_mask_0, x = x_55_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2803_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2804_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2804_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2806_cast_fp16 = sub(x = var_2803_cast_fp16, y = var_2804_cast_fp16)[name = string("op_2806_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2807_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2808_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2808_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2810_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2808_cast_fp16)[name = string("op_2810_cast_fp16")];
            int32 var_2812 = const()[name = string("op_2812"), val = int32(-1)];
            bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_9_cast_fp16 = concat(axis = var_2812, interleave = q_9_interleave_0, values = (var_2806_cast_fp16, var_2810_cast_fp16))[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> first_19_begin_0 = const()[name = string("first_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_19_end_0 = const()[name = string("first_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_19_end_mask_0 = const()[name = string("first_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_19_cast_fp16 = slice_by_index(begin = first_19_begin_0, end = first_19_end_0, end_mask = first_19_end_mask_0, x = x_57_cast_fp16)[name = string("first_19_cast_fp16")];
            tensor<int32, [4]> second_19_begin_0 = const()[name = string("second_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_19_end_0 = const()[name = string("second_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_19_end_mask_0 = const()[name = string("second_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_19_cast_fp16 = slice_by_index(begin = second_19_begin_0, end = second_19_end_0, end_mask = second_19_end_mask_0, x = x_57_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2824_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2824_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2825_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2827_cast_fp16 = sub(x = var_2824_cast_fp16, y = var_2825_cast_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2828_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2829_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2831_cast_fp16 = add(x = var_2828_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2831_cast_fp16")];
            int32 var_2833 = const()[name = string("op_2833"), val = int32(-1)];
            bool k_9_interleave_0 = const()[name = string("k_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_9_cast_fp16 = concat(axis = var_2833, interleave = k_9_interleave_0, values = (var_2827_cast_fp16, var_2831_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2844_begin_0 = const()[name = string("op_2844_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2844_end_0 = const()[name = string("op_2844_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2844_end_mask_0 = const()[name = string("op_2844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2844_stride_0 = const()[name = string("op_2844_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, stride = var_2844_stride_0, x = q_9_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<int32, [4]> var_2854_begin_0 = const()[name = string("op_2854_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2854_end_0 = const()[name = string("op_2854_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2854_end_mask_0 = const()[name = string("op_2854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2854_stride_0 = const()[name = string("op_2854_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2854_cast_fp16 = slice_by_index(begin = var_2854_begin_0, end = var_2854_end_0, end_mask = var_2854_end_mask_0, stride = var_2854_stride_0, x = q_9_cast_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(2)];
            bool stacked_9_interleave_0 = const()[name = string("stacked_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_9_cast_fp16 = concat(axis = var_2856, interleave = stacked_9_interleave_0, values = (var_2844_cast_fp16, var_2854_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2862_begin_0 = const()[name = string("op_2862_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2862_end_0 = const()[name = string("op_2862_end_0"), val = tensor<int32, [4]>([5, 8, 4112, 128])];
            tensor<bool, [4]> var_2862_end_mask_0 = const()[name = string("op_2862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2862_cast_fp16 = slice_by_index(begin = var_2862_begin_0, end = var_2862_end_0, end_mask = var_2862_end_mask_0, x = read_state_0)[name = string("op_2862_cast_fp16")];
            tensor<int32, [4]> var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = var_2862_cast_fp16)[name = string("op_2872_cast_fp16")];
            bool var_2881_transpose_x_1 = const()[name = string("op_2881_transpose_x_1"), val = bool(false)];
            bool var_2881_transpose_y_1 = const()[name = string("op_2881_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_2881_cast_fp16 = matmul(transpose_x = var_2881_transpose_x_1, transpose_y = var_2881_transpose_y_1, x = var_2872_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_2881_cast_fp16")];
            tensor<int32, [4]> var_2884_perm_0 = const()[name = string("op_2884_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2885_to_fp16 = const()[name = string("op_2885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_2884_cast_fp16 = transpose(perm = var_2884_perm_0, x = var_2881_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 4096]> var_2886_cast_fp16 = mul(x = var_2884_cast_fp16, y = var_2885_to_fp16)[name = string("op_2886_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_9_cast_fp16 = add(x = var_2886_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_2892_transpose_x_1 = const()[name = string("op_2892_transpose_x_1"), val = bool(false)];
            bool var_2892_transpose_y_1 = const()[name = string("op_2892_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2892_cast_fp16 = matmul(transpose_x = var_2892_transpose_x_1, transpose_y = var_2892_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_2892_cast_fp16")];
            fp16 var_2893_to_fp16 = const()[name = string("op_2893_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2894_cast_fp16 = mul(x = var_2892_cast_fp16, y = var_2893_to_fp16)[name = string("op_2894_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_2894_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_2898 = const()[name = string("op_2898"), val = int32(-1)];
            bool var_2899_interleave_0 = const()[name = string("op_2899_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_2899_cast_fp16 = concat(axis = var_2898, interleave = var_2899_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_2899_cast_fp16")];
            int32 var_2900 = const()[name = string("op_2900"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_9_cast_fp16 = softmax(axis = var_2900, x = var_2899_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_2907_begin_0 = const()[name = string("op_2907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2907_end_0 = const()[name = string("op_2907_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_2907_end_mask_0 = const()[name = string("op_2907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_2907_cast_fp16 = slice_by_index(begin = var_2907_begin_0, end = var_2907_end_0, end_mask = var_2907_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<int32, [4]> var_2912_begin_0 = const()[name = string("op_2912_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2912_end_0 = const()[name = string("op_2912_end_0"), val = tensor<int32, [4]>([5, 8, 4112, 128])];
            tensor<bool, [4]> var_2912_end_mask_0 = const()[name = string("op_2912_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_2912_cast_fp16 = slice_by_index(begin = var_2912_begin_0, end = var_2912_end_0, end_mask = var_2912_end_mask_0, x = read_state_1)[name = string("op_2912_cast_fp16")];
            tensor<int32, [4]> var_2922_begin_0 = const()[name = string("op_2922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2922_end_0 = const()[name = string("op_2922_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_2922_end_mask_0 = const()[name = string("op_2922_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_2922_cast_fp16 = slice_by_index(begin = var_2922_begin_0, end = var_2922_end_0, end_mask = var_2922_end_mask_0, x = var_2912_cast_fp16)[name = string("op_2922_cast_fp16")];
            bool var_2928_transpose_x_0 = const()[name = string("op_2928_transpose_x_0"), val = bool(false)];
            bool var_2928_transpose_y_0 = const()[name = string("op_2928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2928_cast_fp16 = matmul(transpose_x = var_2928_transpose_x_0, transpose_y = var_2928_transpose_y_0, x = var_2907_cast_fp16, y = var_2922_cast_fp16)[name = string("op_2928_cast_fp16")];
            tensor<int32, [4]> var_2933_begin_0 = const()[name = string("op_2933_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_2933_end_0 = const()[name = string("op_2933_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_2933_end_mask_0 = const()[name = string("op_2933_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2933_cast_fp16 = slice_by_index(begin = var_2933_begin_0, end = var_2933_end_0, end_mask = var_2933_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2933_cast_fp16")];
            bool var_2934_transpose_x_0 = const()[name = string("op_2934_transpose_x_0"), val = bool(false)];
            bool var_2934_transpose_y_0 = const()[name = string("op_2934_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2789_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_2934_cast_fp16 = matmul(transpose_x = var_2934_transpose_x_0, transpose_y = var_2934_transpose_y_0, x = var_2933_cast_fp16, y = v_9_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2934_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_2951_begin_0 = const()[name = string("op_2951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2951_end_0 = const()[name = string("op_2951_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2951_end_mask_0 = const()[name = string("op_2951_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2951_cast_fp16 = slice_by_index(begin = var_2951_begin_0, end = var_2951_end_0, end_mask = var_2951_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2951_cast_fp16")];
            tensor<int32, [4]> var_2966_begin_0 = const()[name = string("op_2966_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2966_end_0 = const()[name = string("op_2966_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2966_end_mask_0 = const()[name = string("op_2966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2966_cast_fp16 = slice_by_index(begin = var_2966_begin_0, end = var_2966_end_0, end_mask = var_2966_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2966_cast_fp16")];
            int32 var_2968 = const()[name = string("op_2968"), val = int32(1)];
            bool mixed_19_interleave_0 = const()[name = string("mixed_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_19_cast_fp16 = concat(axis = var_2968, interleave = mixed_19_interleave_0, values = (var_2951_cast_fp16, var_2966_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_2972_perm_0 = const()[name = string("op_2972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2976 = const()[name = string("op_2976"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2972_cast_fp16 = transpose(perm = var_2972_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_2976, x = var_2972_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378490176))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380587392))))[name = string("model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = linear_31_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2987_promoted_to_fp16 = const()[name = string("op_2987_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2989_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2987_promoted_to_fp16)[name = string("op_2989_cast_fp16")];
            tensor<int32, [1]> var_2991_axes_0 = const()[name = string("op_2991_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2991_keep_dims_0 = const()[name = string("op_2991_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2991_cast_fp16 = reduce_mean(axes = var_2991_axes_0, keep_dims = var_2991_keep_dims_0, x = var_2989_cast_fp16)[name = string("op_2991_cast_fp16")];
            fp16 var_2992_to_fp16 = const()[name = string("op_2992_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2993_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_to_fp16)[name = string("op_2993_cast_fp16")];
            fp32 var_2994_epsilon_0 = const()[name = string("op_2994_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2994_cast_fp16 = rsqrt(epsilon = var_2994_epsilon_0, x = var_2993_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2995_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2994_cast_fp16)[name = string("op_2995_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380589504)))];
            tensor<fp16, [1, 16, 1024]> input_45_cast_fp16 = mul(x = var_2995_cast_fp16, y = model_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380591616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383737408))))[name = string("model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3000_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3000_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383743616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889408))))[name = string("model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_49_cast_fp16 = mul(x = var_3000_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386895616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390041408))))[name = string("model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_34_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_3014_promoted_to_fp16 = const()[name = string("op_3014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3016_cast_fp16 = pow(x = x_61_cast_fp16, y = var_3014_promoted_to_fp16)[name = string("op_3016_cast_fp16")];
            tensor<int32, [1]> var_3018_axes_0 = const()[name = string("op_3018_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3018_keep_dims_0 = const()[name = string("op_3018_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3018_cast_fp16 = reduce_mean(axes = var_3018_axes_0, keep_dims = var_3018_keep_dims_0, x = var_3016_cast_fp16)[name = string("op_3018_cast_fp16")];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3020_cast_fp16 = add(x = var_3018_cast_fp16, y = var_3019_to_fp16)[name = string("op_3020_cast_fp16")];
            fp32 var_3021_epsilon_0 = const()[name = string("op_3021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3021_cast_fp16 = rsqrt(epsilon = var_3021_epsilon_0, x = var_3020_cast_fp16)[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3022_cast_fp16 = mul(x = x_61_cast_fp16, y = var_3021_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390043520)))];
            tensor<fp16, [1, 16, 1024]> input_51_cast_fp16 = mul(x = var_3022_cast_fp16, y = model_layers_5_input_layernorm_weight_to_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390045632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392142848))))[name = string("model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3032_cast_fp16 = reshape(shape = var_3031, x = linear_35_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> x_63_perm_0 = const()[name = string("x_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3040_promoted_to_fp16 = const()[name = string("op_3040_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_63_cast_fp16 = transpose(perm = x_63_perm_0, x = var_3032_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3042_cast_fp16 = pow(x = x_63_cast_fp16, y = var_3040_promoted_to_fp16)[name = string("op_3042_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3042_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 var_3045_to_fp16 = const()[name = string("op_3045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3046_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3045_to_fp16)[name = string("op_3046_cast_fp16")];
            fp32 var_3047_epsilon_0 = const()[name = string("op_3047_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3047_cast_fp16 = rsqrt(epsilon = var_3047_epsilon_0, x = var_3046_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3048_cast_fp16 = mul(x = x_63_cast_fp16, y = var_3047_cast_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147008)))];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = mul(x = var_3048_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393195968))))[name = string("model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3058_cast_fp16 = reshape(shape = var_3057, x = linear_36_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3066_promoted_to_fp16 = const()[name = string("op_3066_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_3058_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3068_cast_fp16 = pow(x = x_65_cast_fp16, y = var_3066_promoted_to_fp16)[name = string("op_3068_cast_fp16")];
            tensor<int32, [1]> var_3070_axes_0 = const()[name = string("op_3070_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3070_keep_dims_0 = const()[name = string("op_3070_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3070_cast_fp16 = reduce_mean(axes = var_3070_axes_0, keep_dims = var_3070_keep_dims_0, x = var_3068_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp16 var_3071_to_fp16 = const()[name = string("op_3071_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3072_cast_fp16 = add(x = var_3070_cast_fp16, y = var_3071_to_fp16)[name = string("op_3072_cast_fp16")];
            fp32 var_3073_epsilon_0 = const()[name = string("op_3073_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3073_cast_fp16 = rsqrt(epsilon = var_3073_epsilon_0, x = var_3072_cast_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3074_cast_fp16 = mul(x = x_65_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198080)))];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = mul(x = var_3074_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394247040))))[name = string("model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3083 = const()[name = string("op_3083"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3084_cast_fp16 = reshape(shape = var_3083, x = linear_37_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_21_begin_0 = const()[name = string("first_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_21_end_0 = const()[name = string("first_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_21_end_mask_0 = const()[name = string("first_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_21_cast_fp16 = slice_by_index(begin = first_21_begin_0, end = first_21_end_0, end_mask = first_21_end_mask_0, x = x_67_cast_fp16)[name = string("first_21_cast_fp16")];
            tensor<int32, [4]> second_21_begin_0 = const()[name = string("second_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_21_end_0 = const()[name = string("second_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_21_end_mask_0 = const()[name = string("second_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_21_cast_fp16 = slice_by_index(begin = second_21_begin_0, end = second_21_end_0, end_mask = second_21_end_mask_0, x = x_67_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3098_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3098_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3099_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3101_cast_fp16 = sub(x = var_3098_cast_fp16, y = var_3099_cast_fp16)[name = string("op_3101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3102_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3102_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3103_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3105_cast_fp16 = add(x = var_3102_cast_fp16, y = var_3103_cast_fp16)[name = string("op_3105_cast_fp16")];
            int32 var_3107 = const()[name = string("op_3107"), val = int32(-1)];
            bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_11_cast_fp16 = concat(axis = var_3107, interleave = q_11_interleave_0, values = (var_3101_cast_fp16, var_3105_cast_fp16))[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> first_23_begin_0 = const()[name = string("first_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_23_end_0 = const()[name = string("first_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_23_end_mask_0 = const()[name = string("first_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_23_cast_fp16 = slice_by_index(begin = first_23_begin_0, end = first_23_end_0, end_mask = first_23_end_mask_0, x = x_69_cast_fp16)[name = string("first_23_cast_fp16")];
            tensor<int32, [4]> second_23_begin_0 = const()[name = string("second_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_23_end_0 = const()[name = string("second_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_23_end_mask_0 = const()[name = string("second_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_23_cast_fp16 = slice_by_index(begin = second_23_begin_0, end = second_23_end_0, end_mask = second_23_end_mask_0, x = x_69_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3119_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3119_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3120_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3122_cast_fp16 = sub(x = var_3119_cast_fp16, y = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3123_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3123_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3124_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3124_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3126_cast_fp16 = add(x = var_3123_cast_fp16, y = var_3124_cast_fp16)[name = string("op_3126_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-1)];
            bool k_11_interleave_0 = const()[name = string("k_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_11_cast_fp16 = concat(axis = var_3128, interleave = k_11_interleave_0, values = (var_3122_cast_fp16, var_3126_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3139_begin_0 = const()[name = string("op_3139_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3139_end_0 = const()[name = string("op_3139_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3139_end_mask_0 = const()[name = string("op_3139_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3139_stride_0 = const()[name = string("op_3139_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3139_cast_fp16 = slice_by_index(begin = var_3139_begin_0, end = var_3139_end_0, end_mask = var_3139_end_mask_0, stride = var_3139_stride_0, x = q_11_cast_fp16)[name = string("op_3139_cast_fp16")];
            tensor<int32, [4]> var_3149_begin_0 = const()[name = string("op_3149_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3149_end_0 = const()[name = string("op_3149_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3149_end_mask_0 = const()[name = string("op_3149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3149_stride_0 = const()[name = string("op_3149_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3149_cast_fp16 = slice_by_index(begin = var_3149_begin_0, end = var_3149_end_0, end_mask = var_3149_end_mask_0, stride = var_3149_stride_0, x = q_11_cast_fp16)[name = string("op_3149_cast_fp16")];
            int32 var_3151 = const()[name = string("op_3151"), val = int32(2)];
            bool stacked_11_interleave_0 = const()[name = string("stacked_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_11_cast_fp16 = concat(axis = var_3151, interleave = stacked_11_interleave_0, values = (var_3139_cast_fp16, var_3149_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3157_begin_0 = const()[name = string("op_3157_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3157_end_0 = const()[name = string("op_3157_end_0"), val = tensor<int32, [4]>([6, 8, 4112, 128])];
            tensor<bool, [4]> var_3157_end_mask_0 = const()[name = string("op_3157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3157_cast_fp16 = slice_by_index(begin = var_3157_begin_0, end = var_3157_end_0, end_mask = var_3157_end_mask_0, x = read_state_0)[name = string("op_3157_cast_fp16")];
            tensor<int32, [4]> var_3167_begin_0 = const()[name = string("op_3167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3167_end_0 = const()[name = string("op_3167_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3167_end_mask_0 = const()[name = string("op_3167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, x = var_3157_cast_fp16)[name = string("op_3167_cast_fp16")];
            bool var_3176_transpose_x_1 = const()[name = string("op_3176_transpose_x_1"), val = bool(false)];
            bool var_3176_transpose_y_1 = const()[name = string("op_3176_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3176_cast_fp16 = matmul(transpose_x = var_3176_transpose_x_1, transpose_y = var_3176_transpose_y_1, x = var_3167_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3176_cast_fp16")];
            tensor<int32, [4]> var_3179_perm_0 = const()[name = string("op_3179_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3180_to_fp16 = const()[name = string("op_3180_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3179_cast_fp16 = transpose(perm = var_3179_perm_0, x = var_3176_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 4096]> var_3181_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_3180_to_fp16)[name = string("op_3181_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_11_cast_fp16 = add(x = var_3181_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3187_transpose_x_1 = const()[name = string("op_3187_transpose_x_1"), val = bool(false)];
            bool var_3187_transpose_y_1 = const()[name = string("op_3187_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3187_cast_fp16 = matmul(transpose_x = var_3187_transpose_x_1, transpose_y = var_3187_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3187_cast_fp16")];
            fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3189_cast_fp16 = mul(x = var_3187_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3189_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3193 = const()[name = string("op_3193"), val = int32(-1)];
            bool var_3194_interleave_0 = const()[name = string("op_3194_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3194_cast_fp16 = concat(axis = var_3193, interleave = var_3194_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3194_cast_fp16")];
            int32 var_3195 = const()[name = string("op_3195"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_11_cast_fp16 = softmax(axis = var_3195, x = var_3194_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3202_begin_0 = const()[name = string("op_3202_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3202_end_0 = const()[name = string("op_3202_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3202_end_mask_0 = const()[name = string("op_3202_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3202_cast_fp16 = slice_by_index(begin = var_3202_begin_0, end = var_3202_end_0, end_mask = var_3202_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3202_cast_fp16")];
            tensor<int32, [4]> var_3207_begin_0 = const()[name = string("op_3207_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3207_end_0 = const()[name = string("op_3207_end_0"), val = tensor<int32, [4]>([6, 8, 4112, 128])];
            tensor<bool, [4]> var_3207_end_mask_0 = const()[name = string("op_3207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3207_cast_fp16 = slice_by_index(begin = var_3207_begin_0, end = var_3207_end_0, end_mask = var_3207_end_mask_0, x = read_state_1)[name = string("op_3207_cast_fp16")];
            tensor<int32, [4]> var_3217_begin_0 = const()[name = string("op_3217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3217_end_0 = const()[name = string("op_3217_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3217_end_mask_0 = const()[name = string("op_3217_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3217_cast_fp16 = slice_by_index(begin = var_3217_begin_0, end = var_3217_end_0, end_mask = var_3217_end_mask_0, x = var_3207_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3223_transpose_x_0 = const()[name = string("op_3223_transpose_x_0"), val = bool(false)];
            bool var_3223_transpose_y_0 = const()[name = string("op_3223_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3223_cast_fp16 = matmul(transpose_x = var_3223_transpose_x_0, transpose_y = var_3223_transpose_y_0, x = var_3202_cast_fp16, y = var_3217_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<int32, [4]> var_3228_begin_0 = const()[name = string("op_3228_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3228_end_0 = const()[name = string("op_3228_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3228_end_mask_0 = const()[name = string("op_3228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3228_cast_fp16 = slice_by_index(begin = var_3228_begin_0, end = var_3228_end_0, end_mask = var_3228_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3229_transpose_x_0 = const()[name = string("op_3229_transpose_x_0"), val = bool(false)];
            bool var_3229_transpose_y_0 = const()[name = string("op_3229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3084_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3229_cast_fp16 = matmul(transpose_x = var_3229_transpose_x_0, transpose_y = var_3229_transpose_y_0, x = var_3228_cast_fp16, y = v_11_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3246_begin_0 = const()[name = string("op_3246_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3246_end_0 = const()[name = string("op_3246_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3246_end_mask_0 = const()[name = string("op_3246_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3246_cast_fp16 = slice_by_index(begin = var_3246_begin_0, end = var_3246_end_0, end_mask = var_3246_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3246_cast_fp16")];
            tensor<int32, [4]> var_3261_begin_0 = const()[name = string("op_3261_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3261_end_0 = const()[name = string("op_3261_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3261_end_mask_0 = const()[name = string("op_3261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3261_cast_fp16 = slice_by_index(begin = var_3261_begin_0, end = var_3261_end_0, end_mask = var_3261_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3261_cast_fp16")];
            int32 var_3263 = const()[name = string("op_3263"), val = int32(1)];
            bool mixed_23_interleave_0 = const()[name = string("mixed_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_23_cast_fp16 = concat(axis = var_3263, interleave = mixed_23_interleave_0, values = (var_3246_cast_fp16, var_3261_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3267_perm_0 = const()[name = string("op_3267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3267_cast_fp16 = transpose(perm = var_3267_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3271, x = var_3267_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394249152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396346368))))[name = string("model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_71_cast_fp16 = add(x = x_61_cast_fp16, y = linear_38_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_3282_promoted_to_fp16 = const()[name = string("op_3282_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3284_cast_fp16 = pow(x = x_71_cast_fp16, y = var_3282_promoted_to_fp16)[name = string("op_3284_cast_fp16")];
            tensor<int32, [1]> var_3286_axes_0 = const()[name = string("op_3286_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3286_keep_dims_0 = const()[name = string("op_3286_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3286_cast_fp16 = reduce_mean(axes = var_3286_axes_0, keep_dims = var_3286_keep_dims_0, x = var_3284_cast_fp16)[name = string("op_3286_cast_fp16")];
            fp16 var_3287_to_fp16 = const()[name = string("op_3287_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3288_cast_fp16 = add(x = var_3286_cast_fp16, y = var_3287_to_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3289_cast_fp16 = rsqrt(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3290_cast_fp16 = mul(x = x_71_cast_fp16, y = var_3289_cast_fp16)[name = string("op_3290_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396348480)))];
            tensor<fp16, [1, 16, 1024]> input_55_cast_fp16 = mul(x = var_3290_cast_fp16, y = model_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396350592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399496384))))[name = string("model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3295_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3295_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399502592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402648384))))[name = string("model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_59_cast_fp16 = mul(x = var_3295_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402654592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405800384))))[name = string("model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_41_cast_fp16)[name = string("x_73_cast_fp16")];
            fp16 var_3309_promoted_to_fp16 = const()[name = string("op_3309_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3311_cast_fp16 = pow(x = x_73_cast_fp16, y = var_3309_promoted_to_fp16)[name = string("op_3311_cast_fp16")];
            tensor<int32, [1]> var_3313_axes_0 = const()[name = string("op_3313_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3313_keep_dims_0 = const()[name = string("op_3313_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3313_cast_fp16 = reduce_mean(axes = var_3313_axes_0, keep_dims = var_3313_keep_dims_0, x = var_3311_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3315_cast_fp16 = add(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            fp32 var_3316_epsilon_0 = const()[name = string("op_3316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3316_cast_fp16 = rsqrt(epsilon = var_3316_epsilon_0, x = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3317_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405802496)))];
            tensor<fp16, [1, 16, 1024]> input_61_cast_fp16 = mul(x = var_3317_cast_fp16, y = model_layers_6_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405804608))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407901824))))[name = string("model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_3326 = const()[name = string("op_3326"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3327_cast_fp16 = reshape(shape = var_3326, x = linear_42_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3335_promoted_to_fp16 = const()[name = string("op_3335_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3327_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3337_cast_fp16 = pow(x = x_75_cast_fp16, y = var_3335_promoted_to_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]>([-1])];
            bool var_3339_keep_dims_0 = const()[name = string("op_3339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3339_cast_fp16 = reduce_mean(axes = var_3339_axes_0, keep_dims = var_3339_keep_dims_0, x = var_3337_cast_fp16)[name = string("op_3339_cast_fp16")];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3341_cast_fp16 = add(x = var_3339_cast_fp16, y = var_3340_to_fp16)[name = string("op_3341_cast_fp16")];
            fp32 var_3342_epsilon_0 = const()[name = string("op_3342_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3342_cast_fp16 = rsqrt(epsilon = var_3342_epsilon_0, x = var_3341_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3343_cast_fp16 = mul(x = x_75_cast_fp16, y = var_3342_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407905984)))];
            tensor<fp16, [1, 16, 16, 128]> x_79_cast_fp16 = mul(x = var_3343_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407906304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408954944))))[name = string("model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3352 = const()[name = string("op_3352"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3353_cast_fp16 = reshape(shape = var_3352, x = linear_43_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3361_promoted_to_fp16 = const()[name = string("op_3361_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3353_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3363_cast_fp16 = pow(x = x_77_cast_fp16, y = var_3361_promoted_to_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3365_keep_dims_0 = const()[name = string("op_3365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3365_cast_fp16 = reduce_mean(axes = var_3365_axes_0, keep_dims = var_3365_keep_dims_0, x = var_3363_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3367_cast_fp16 = add(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("op_3367_cast_fp16")];
            fp32 var_3368_epsilon_0 = const()[name = string("op_3368_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3368_cast_fp16 = rsqrt(epsilon = var_3368_epsilon_0, x = var_3367_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3369_cast_fp16 = mul(x = x_77_cast_fp16, y = var_3368_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957056)))];
            tensor<fp16, [1, 8, 16, 128]> x_81_cast_fp16 = mul(x = var_3369_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410006016))))[name = string("model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3378 = const()[name = string("op_3378"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3379_cast_fp16 = reshape(shape = var_3378, x = linear_44_cast_fp16)[name = string("op_3379_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_25_begin_0 = const()[name = string("first_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_25_end_0 = const()[name = string("first_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_25_end_mask_0 = const()[name = string("first_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_25_cast_fp16 = slice_by_index(begin = first_25_begin_0, end = first_25_end_0, end_mask = first_25_end_mask_0, x = x_79_cast_fp16)[name = string("first_25_cast_fp16")];
            tensor<int32, [4]> second_25_begin_0 = const()[name = string("second_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_25_end_0 = const()[name = string("second_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_25_end_mask_0 = const()[name = string("second_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_25_cast_fp16 = slice_by_index(begin = second_25_begin_0, end = second_25_end_0, end_mask = second_25_end_mask_0, x = x_79_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3393_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3393_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3394_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3396_cast_fp16 = sub(x = var_3393_cast_fp16, y = var_3394_cast_fp16)[name = string("op_3396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3397_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3398_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_13_cast_fp16 = concat(axis = var_3402, interleave = q_13_interleave_0, values = (var_3396_cast_fp16, var_3400_cast_fp16))[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> first_27_begin_0 = const()[name = string("first_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_27_end_0 = const()[name = string("first_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_27_end_mask_0 = const()[name = string("first_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_27_cast_fp16 = slice_by_index(begin = first_27_begin_0, end = first_27_end_0, end_mask = first_27_end_mask_0, x = x_81_cast_fp16)[name = string("first_27_cast_fp16")];
            tensor<int32, [4]> second_27_begin_0 = const()[name = string("second_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_27_end_0 = const()[name = string("second_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_27_end_mask_0 = const()[name = string("second_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_27_cast_fp16 = slice_by_index(begin = second_27_begin_0, end = second_27_end_0, end_mask = second_27_end_mask_0, x = x_81_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3414_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3414_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3415_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3415_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3417_cast_fp16 = sub(x = var_3414_cast_fp16, y = var_3415_cast_fp16)[name = string("op_3417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3418_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3419_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3421_cast_fp16 = add(x = var_3418_cast_fp16, y = var_3419_cast_fp16)[name = string("op_3421_cast_fp16")];
            int32 var_3423 = const()[name = string("op_3423"), val = int32(-1)];
            bool k_13_interleave_0 = const()[name = string("k_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_13_cast_fp16 = concat(axis = var_3423, interleave = k_13_interleave_0, values = (var_3417_cast_fp16, var_3421_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3434_begin_0 = const()[name = string("op_3434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3434_end_0 = const()[name = string("op_3434_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3434_end_mask_0 = const()[name = string("op_3434_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3434_stride_0 = const()[name = string("op_3434_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3434_cast_fp16 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, stride = var_3434_stride_0, x = q_13_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<int32, [4]> var_3444_begin_0 = const()[name = string("op_3444_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3444_end_0 = const()[name = string("op_3444_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3444_end_mask_0 = const()[name = string("op_3444_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3444_stride_0 = const()[name = string("op_3444_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3444_cast_fp16 = slice_by_index(begin = var_3444_begin_0, end = var_3444_end_0, end_mask = var_3444_end_mask_0, stride = var_3444_stride_0, x = q_13_cast_fp16)[name = string("op_3444_cast_fp16")];
            int32 var_3446 = const()[name = string("op_3446"), val = int32(2)];
            bool stacked_13_interleave_0 = const()[name = string("stacked_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_13_cast_fp16 = concat(axis = var_3446, interleave = stacked_13_interleave_0, values = (var_3434_cast_fp16, var_3444_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3452_begin_0 = const()[name = string("op_3452_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3452_end_0 = const()[name = string("op_3452_end_0"), val = tensor<int32, [4]>([7, 8, 4112, 128])];
            tensor<bool, [4]> var_3452_end_mask_0 = const()[name = string("op_3452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3452_cast_fp16 = slice_by_index(begin = var_3452_begin_0, end = var_3452_end_0, end_mask = var_3452_end_mask_0, x = read_state_0)[name = string("op_3452_cast_fp16")];
            tensor<int32, [4]> var_3462_begin_0 = const()[name = string("op_3462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3462_end_0 = const()[name = string("op_3462_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3462_end_mask_0 = const()[name = string("op_3462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3462_cast_fp16 = slice_by_index(begin = var_3462_begin_0, end = var_3462_end_0, end_mask = var_3462_end_mask_0, x = var_3452_cast_fp16)[name = string("op_3462_cast_fp16")];
            bool var_3471_transpose_x_1 = const()[name = string("op_3471_transpose_x_1"), val = bool(false)];
            bool var_3471_transpose_y_1 = const()[name = string("op_3471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3471_cast_fp16 = matmul(transpose_x = var_3471_transpose_x_1, transpose_y = var_3471_transpose_y_1, x = var_3462_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<int32, [4]> var_3474_perm_0 = const()[name = string("op_3474_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3475_to_fp16 = const()[name = string("op_3475_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3474_cast_fp16 = transpose(perm = var_3474_perm_0, x = var_3471_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 4096]> var_3476_cast_fp16 = mul(x = var_3474_cast_fp16, y = var_3475_to_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_13_cast_fp16 = add(x = var_3476_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3482_transpose_x_1 = const()[name = string("op_3482_transpose_x_1"), val = bool(false)];
            bool var_3482_transpose_y_1 = const()[name = string("op_3482_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3482_cast_fp16 = matmul(transpose_x = var_3482_transpose_x_1, transpose_y = var_3482_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3482_cast_fp16")];
            fp16 var_3483_to_fp16 = const()[name = string("op_3483_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3484_cast_fp16 = mul(x = var_3482_cast_fp16, y = var_3483_to_fp16)[name = string("op_3484_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3484_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3488 = const()[name = string("op_3488"), val = int32(-1)];
            bool var_3489_interleave_0 = const()[name = string("op_3489_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3489_cast_fp16 = concat(axis = var_3488, interleave = var_3489_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3489_cast_fp16")];
            int32 var_3490 = const()[name = string("op_3490"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_13_cast_fp16 = softmax(axis = var_3490, x = var_3489_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3497_begin_0 = const()[name = string("op_3497_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3497_end_0 = const()[name = string("op_3497_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3497_end_mask_0 = const()[name = string("op_3497_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3497_cast_fp16 = slice_by_index(begin = var_3497_begin_0, end = var_3497_end_0, end_mask = var_3497_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> var_3502_begin_0 = const()[name = string("op_3502_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3502_end_0 = const()[name = string("op_3502_end_0"), val = tensor<int32, [4]>([7, 8, 4112, 128])];
            tensor<bool, [4]> var_3502_end_mask_0 = const()[name = string("op_3502_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3502_cast_fp16 = slice_by_index(begin = var_3502_begin_0, end = var_3502_end_0, end_mask = var_3502_end_mask_0, x = read_state_1)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> var_3512_begin_0 = const()[name = string("op_3512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3512_end_0 = const()[name = string("op_3512_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3512_end_mask_0 = const()[name = string("op_3512_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3512_cast_fp16 = slice_by_index(begin = var_3512_begin_0, end = var_3512_end_0, end_mask = var_3512_end_mask_0, x = var_3502_cast_fp16)[name = string("op_3512_cast_fp16")];
            bool var_3518_transpose_x_0 = const()[name = string("op_3518_transpose_x_0"), val = bool(false)];
            bool var_3518_transpose_y_0 = const()[name = string("op_3518_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3518_cast_fp16 = matmul(transpose_x = var_3518_transpose_x_0, transpose_y = var_3518_transpose_y_0, x = var_3497_cast_fp16, y = var_3512_cast_fp16)[name = string("op_3518_cast_fp16")];
            tensor<int32, [4]> var_3523_begin_0 = const()[name = string("op_3523_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3523_end_0 = const()[name = string("op_3523_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3523_end_mask_0 = const()[name = string("op_3523_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3523_cast_fp16 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3523_cast_fp16")];
            bool var_3524_transpose_x_0 = const()[name = string("op_3524_transpose_x_0"), val = bool(false)];
            bool var_3524_transpose_y_0 = const()[name = string("op_3524_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3379_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3524_cast_fp16 = matmul(transpose_x = var_3524_transpose_x_0, transpose_y = var_3524_transpose_y_0, x = var_3523_cast_fp16, y = v_13_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3524_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3541_begin_0 = const()[name = string("op_3541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3541_end_0 = const()[name = string("op_3541_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3541_end_mask_0 = const()[name = string("op_3541_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3541_cast_fp16 = slice_by_index(begin = var_3541_begin_0, end = var_3541_end_0, end_mask = var_3541_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3541_cast_fp16")];
            tensor<int32, [4]> var_3556_begin_0 = const()[name = string("op_3556_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3556_end_0 = const()[name = string("op_3556_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3556_end_mask_0 = const()[name = string("op_3556_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3556_cast_fp16 = slice_by_index(begin = var_3556_begin_0, end = var_3556_end_0, end_mask = var_3556_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3556_cast_fp16")];
            int32 var_3558 = const()[name = string("op_3558"), val = int32(1)];
            bool mixed_27_interleave_0 = const()[name = string("mixed_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_27_cast_fp16 = concat(axis = var_3558, interleave = mixed_27_interleave_0, values = (var_3541_cast_fp16, var_3556_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3562_perm_0 = const()[name = string("op_3562_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3562_cast_fp16 = transpose(perm = var_3562_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3566, x = var_3562_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410008128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412105344))))[name = string("model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = linear_45_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_3577_promoted_to_fp16 = const()[name = string("op_3577_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3579_cast_fp16 = pow(x = x_83_cast_fp16, y = var_3577_promoted_to_fp16)[name = string("op_3579_cast_fp16")];
            tensor<int32, [1]> var_3581_axes_0 = const()[name = string("op_3581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3581_keep_dims_0 = const()[name = string("op_3581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3581_cast_fp16 = reduce_mean(axes = var_3581_axes_0, keep_dims = var_3581_keep_dims_0, x = var_3579_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp16 var_3582_to_fp16 = const()[name = string("op_3582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3583_cast_fp16 = add(x = var_3581_cast_fp16, y = var_3582_to_fp16)[name = string("op_3583_cast_fp16")];
            fp32 var_3584_epsilon_0 = const()[name = string("op_3584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3584_cast_fp16 = rsqrt(epsilon = var_3584_epsilon_0, x = var_3583_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3585_cast_fp16 = mul(x = x_83_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412107456)))];
            tensor<fp16, [1, 16, 1024]> input_65_cast_fp16 = mul(x = var_3585_cast_fp16, y = model_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412109568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415255360))))[name = string("model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3590_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3590_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415261568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418407360))))[name = string("model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_69_cast_fp16 = mul(x = var_3590_cast_fp16, y = linear_47_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418413568))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421559360))))[name = string("model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_48_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_3604_promoted_to_fp16 = const()[name = string("op_3604_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3606_cast_fp16 = pow(x = x_85_cast_fp16, y = var_3604_promoted_to_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3608_keep_dims_0 = const()[name = string("op_3608_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3608_cast_fp16 = reduce_mean(axes = var_3608_axes_0, keep_dims = var_3608_keep_dims_0, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3610_cast_fp16 = add(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("op_3610_cast_fp16")];
            fp32 var_3611_epsilon_0 = const()[name = string("op_3611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3611_cast_fp16 = rsqrt(epsilon = var_3611_epsilon_0, x = var_3610_cast_fp16)[name = string("op_3611_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3612_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3611_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421561472)))];
            tensor<fp16, [1, 16, 1024]> input_71_cast_fp16 = mul(x = var_3612_cast_fp16, y = model_layers_7_input_layernorm_weight_to_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421563584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423660800))))[name = string("model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3622_cast_fp16 = reshape(shape = var_3621, x = linear_49_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3630_promoted_to_fp16 = const()[name = string("op_3630_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_3622_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3632_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3630_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [1]> var_3634_axes_0 = const()[name = string("op_3634_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3634_keep_dims_0 = const()[name = string("op_3634_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3634_cast_fp16 = reduce_mean(axes = var_3634_axes_0, keep_dims = var_3634_keep_dims_0, x = var_3632_cast_fp16)[name = string("op_3634_cast_fp16")];
            fp16 var_3635_to_fp16 = const()[name = string("op_3635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3636_cast_fp16 = add(x = var_3634_cast_fp16, y = var_3635_to_fp16)[name = string("op_3636_cast_fp16")];
            fp32 var_3637_epsilon_0 = const()[name = string("op_3637_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3637_cast_fp16 = rsqrt(epsilon = var_3637_epsilon_0, x = var_3636_cast_fp16)[name = string("op_3637_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3638_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3637_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423664960)))];
            tensor<fp16, [1, 16, 16, 128]> x_91_cast_fp16 = mul(x = var_3638_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423665280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424713920))))[name = string("model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3647 = const()[name = string("op_3647"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3648_cast_fp16 = reshape(shape = var_3647, x = linear_50_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3656_promoted_to_fp16 = const()[name = string("op_3656_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_3648_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3658_cast_fp16 = pow(x = x_89_cast_fp16, y = var_3656_promoted_to_fp16)[name = string("op_3658_cast_fp16")];
            tensor<int32, [1]> var_3660_axes_0 = const()[name = string("op_3660_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3660_keep_dims_0 = const()[name = string("op_3660_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3660_cast_fp16 = reduce_mean(axes = var_3660_axes_0, keep_dims = var_3660_keep_dims_0, x = var_3658_cast_fp16)[name = string("op_3660_cast_fp16")];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3662_cast_fp16 = add(x = var_3660_cast_fp16, y = var_3661_to_fp16)[name = string("op_3662_cast_fp16")];
            fp32 var_3663_epsilon_0 = const()[name = string("op_3663_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3663_cast_fp16 = rsqrt(epsilon = var_3663_epsilon_0, x = var_3662_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3664_cast_fp16 = mul(x = x_89_cast_fp16, y = var_3663_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716032)))];
            tensor<fp16, [1, 8, 16, 128]> x_93_cast_fp16 = mul(x = var_3664_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425764992))))[name = string("model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3673 = const()[name = string("op_3673"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3674_cast_fp16 = reshape(shape = var_3673, x = linear_51_cast_fp16)[name = string("op_3674_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_29_begin_0 = const()[name = string("first_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_29_end_0 = const()[name = string("first_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_29_end_mask_0 = const()[name = string("first_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_29_cast_fp16 = slice_by_index(begin = first_29_begin_0, end = first_29_end_0, end_mask = first_29_end_mask_0, x = x_91_cast_fp16)[name = string("first_29_cast_fp16")];
            tensor<int32, [4]> second_29_begin_0 = const()[name = string("second_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_29_end_0 = const()[name = string("second_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_29_end_mask_0 = const()[name = string("second_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_29_cast_fp16 = slice_by_index(begin = second_29_begin_0, end = second_29_end_0, end_mask = second_29_end_mask_0, x = x_91_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3688_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3689_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3691_cast_fp16 = sub(x = var_3688_cast_fp16, y = var_3689_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3692_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3693_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3695_cast_fp16 = add(x = var_3692_cast_fp16, y = var_3693_cast_fp16)[name = string("op_3695_cast_fp16")];
            int32 var_3697 = const()[name = string("op_3697"), val = int32(-1)];
            bool q_15_interleave_0 = const()[name = string("q_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_15_cast_fp16 = concat(axis = var_3697, interleave = q_15_interleave_0, values = (var_3691_cast_fp16, var_3695_cast_fp16))[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> first_31_begin_0 = const()[name = string("first_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_31_end_0 = const()[name = string("first_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_31_end_mask_0 = const()[name = string("first_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_31_cast_fp16 = slice_by_index(begin = first_31_begin_0, end = first_31_end_0, end_mask = first_31_end_mask_0, x = x_93_cast_fp16)[name = string("first_31_cast_fp16")];
            tensor<int32, [4]> second_31_begin_0 = const()[name = string("second_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_31_end_0 = const()[name = string("second_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_31_end_mask_0 = const()[name = string("second_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_31_cast_fp16 = slice_by_index(begin = second_31_begin_0, end = second_31_end_0, end_mask = second_31_end_mask_0, x = x_93_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3709_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3709_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3710_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3710_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3712_cast_fp16 = sub(x = var_3709_cast_fp16, y = var_3710_cast_fp16)[name = string("op_3712_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3713_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3714_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3714_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3716_cast_fp16 = add(x = var_3713_cast_fp16, y = var_3714_cast_fp16)[name = string("op_3716_cast_fp16")];
            int32 var_3718 = const()[name = string("op_3718"), val = int32(-1)];
            bool k_15_interleave_0 = const()[name = string("k_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_15_cast_fp16 = concat(axis = var_3718, interleave = k_15_interleave_0, values = (var_3712_cast_fp16, var_3716_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3729_begin_0 = const()[name = string("op_3729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3729_end_0 = const()[name = string("op_3729_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3729_end_mask_0 = const()[name = string("op_3729_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3729_stride_0 = const()[name = string("op_3729_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3729_cast_fp16 = slice_by_index(begin = var_3729_begin_0, end = var_3729_end_0, end_mask = var_3729_end_mask_0, stride = var_3729_stride_0, x = q_15_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [4]> var_3739_begin_0 = const()[name = string("op_3739_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3739_end_0 = const()[name = string("op_3739_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3739_end_mask_0 = const()[name = string("op_3739_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3739_stride_0 = const()[name = string("op_3739_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3739_cast_fp16 = slice_by_index(begin = var_3739_begin_0, end = var_3739_end_0, end_mask = var_3739_end_mask_0, stride = var_3739_stride_0, x = q_15_cast_fp16)[name = string("op_3739_cast_fp16")];
            int32 var_3741 = const()[name = string("op_3741"), val = int32(2)];
            bool stacked_15_interleave_0 = const()[name = string("stacked_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_15_cast_fp16 = concat(axis = var_3741, interleave = stacked_15_interleave_0, values = (var_3729_cast_fp16, var_3739_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3747_begin_0 = const()[name = string("op_3747_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3747_end_0 = const()[name = string("op_3747_end_0"), val = tensor<int32, [4]>([8, 8, 4112, 128])];
            tensor<bool, [4]> var_3747_end_mask_0 = const()[name = string("op_3747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3747_cast_fp16 = slice_by_index(begin = var_3747_begin_0, end = var_3747_end_0, end_mask = var_3747_end_mask_0, x = read_state_0)[name = string("op_3747_cast_fp16")];
            tensor<int32, [4]> var_3757_begin_0 = const()[name = string("op_3757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3757_end_0 = const()[name = string("op_3757_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3757_end_mask_0 = const()[name = string("op_3757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3757_cast_fp16 = slice_by_index(begin = var_3757_begin_0, end = var_3757_end_0, end_mask = var_3757_end_mask_0, x = var_3747_cast_fp16)[name = string("op_3757_cast_fp16")];
            bool var_3766_transpose_x_1 = const()[name = string("op_3766_transpose_x_1"), val = bool(false)];
            bool var_3766_transpose_y_1 = const()[name = string("op_3766_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_3766_cast_fp16 = matmul(transpose_x = var_3766_transpose_x_1, transpose_y = var_3766_transpose_y_1, x = var_3757_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<int32, [4]> var_3769_perm_0 = const()[name = string("op_3769_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3770_to_fp16 = const()[name = string("op_3770_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_3769_cast_fp16 = transpose(perm = var_3769_perm_0, x = var_3766_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 4096]> var_3771_cast_fp16 = mul(x = var_3769_cast_fp16, y = var_3770_to_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_15_cast_fp16 = add(x = var_3771_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3777_transpose_x_1 = const()[name = string("op_3777_transpose_x_1"), val = bool(false)];
            bool var_3777_transpose_y_1 = const()[name = string("op_3777_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3777_cast_fp16 = matmul(transpose_x = var_3777_transpose_x_1, transpose_y = var_3777_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3779_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_3783 = const()[name = string("op_3783"), val = int32(-1)];
            bool var_3784_interleave_0 = const()[name = string("op_3784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_3784_cast_fp16 = concat(axis = var_3783, interleave = var_3784_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_3784_cast_fp16")];
            int32 var_3785 = const()[name = string("op_3785"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_15_cast_fp16 = softmax(axis = var_3785, x = var_3784_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_3792_begin_0 = const()[name = string("op_3792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3792_end_0 = const()[name = string("op_3792_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_3792_end_mask_0 = const()[name = string("op_3792_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<int32, [4]> var_3797_begin_0 = const()[name = string("op_3797_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3797_end_0 = const()[name = string("op_3797_end_0"), val = tensor<int32, [4]>([8, 8, 4112, 128])];
            tensor<bool, [4]> var_3797_end_mask_0 = const()[name = string("op_3797_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_3797_cast_fp16 = slice_by_index(begin = var_3797_begin_0, end = var_3797_end_0, end_mask = var_3797_end_mask_0, x = read_state_1)[name = string("op_3797_cast_fp16")];
            tensor<int32, [4]> var_3807_begin_0 = const()[name = string("op_3807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3807_end_0 = const()[name = string("op_3807_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_3807_end_mask_0 = const()[name = string("op_3807_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_3807_cast_fp16 = slice_by_index(begin = var_3807_begin_0, end = var_3807_end_0, end_mask = var_3807_end_mask_0, x = var_3797_cast_fp16)[name = string("op_3807_cast_fp16")];
            bool var_3813_transpose_x_0 = const()[name = string("op_3813_transpose_x_0"), val = bool(false)];
            bool var_3813_transpose_y_0 = const()[name = string("op_3813_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3813_cast_fp16 = matmul(transpose_x = var_3813_transpose_x_0, transpose_y = var_3813_transpose_y_0, x = var_3792_cast_fp16, y = var_3807_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [4]> var_3818_begin_0 = const()[name = string("op_3818_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_3818_end_0 = const()[name = string("op_3818_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_3818_end_mask_0 = const()[name = string("op_3818_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3818_cast_fp16 = slice_by_index(begin = var_3818_begin_0, end = var_3818_end_0, end_mask = var_3818_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3818_cast_fp16")];
            bool var_3819_transpose_x_0 = const()[name = string("op_3819_transpose_x_0"), val = bool(false)];
            bool var_3819_transpose_y_0 = const()[name = string("op_3819_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3674_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_3819_cast_fp16 = matmul(transpose_x = var_3819_transpose_x_0, transpose_y = var_3819_transpose_y_0, x = var_3818_cast_fp16, y = v_15_cast_fp16)[name = string("op_3819_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_3813_cast_fp16, y = var_3819_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_3836_begin_0 = const()[name = string("op_3836_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3836_end_0 = const()[name = string("op_3836_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3836_end_mask_0 = const()[name = string("op_3836_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<int32, [4]> var_3851_begin_0 = const()[name = string("op_3851_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3851_end_0 = const()[name = string("op_3851_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3851_end_mask_0 = const()[name = string("op_3851_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3851_cast_fp16 = slice_by_index(begin = var_3851_begin_0, end = var_3851_end_0, end_mask = var_3851_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3851_cast_fp16")];
            int32 var_3853 = const()[name = string("op_3853"), val = int32(1)];
            bool mixed_31_interleave_0 = const()[name = string("mixed_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_31_cast_fp16 = concat(axis = var_3853, interleave = mixed_31_interleave_0, values = (var_3836_cast_fp16, var_3851_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_3857_perm_0 = const()[name = string("op_3857_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3857_cast_fp16 = transpose(perm = var_3857_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425767104))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427864320))))[name = string("model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_95_cast_fp16 = add(x = x_85_cast_fp16, y = linear_52_cast_fp16)[name = string("x_95_cast_fp16")];
            fp16 var_3872_promoted_to_fp16 = const()[name = string("op_3872_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3874_cast_fp16 = pow(x = x_95_cast_fp16, y = var_3872_promoted_to_fp16)[name = string("op_3874_cast_fp16")];
            tensor<int32, [1]> var_3876_axes_0 = const()[name = string("op_3876_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3876_keep_dims_0 = const()[name = string("op_3876_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3876_cast_fp16 = reduce_mean(axes = var_3876_axes_0, keep_dims = var_3876_keep_dims_0, x = var_3874_cast_fp16)[name = string("op_3876_cast_fp16")];
            fp16 var_3877_to_fp16 = const()[name = string("op_3877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3878_cast_fp16 = add(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            fp32 var_3879_epsilon_0 = const()[name = string("op_3879_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3879_cast_fp16 = rsqrt(epsilon = var_3879_epsilon_0, x = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3880_cast_fp16 = mul(x = x_95_cast_fp16, y = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427866432)))];
            tensor<fp16, [1, 16, 1024]> input_75_cast_fp16 = mul(x = var_3880_cast_fp16, y = model_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427868544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431014336))))[name = string("model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3885_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_3885_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431020544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434166336))))[name = string("model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_79_cast_fp16 = mul(x = var_3885_cast_fp16, y = linear_54_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434172544))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437318336))))[name = string("model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_97_cast_fp16 = add(x = x_95_cast_fp16, y = linear_55_cast_fp16)[name = string("x_97_cast_fp16")];
            fp16 var_3899_promoted_to_fp16 = const()[name = string("op_3899_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3901_cast_fp16 = pow(x = x_97_cast_fp16, y = var_3899_promoted_to_fp16)[name = string("op_3901_cast_fp16")];
            tensor<int32, [1]> var_3903_axes_0 = const()[name = string("op_3903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3903_keep_dims_0 = const()[name = string("op_3903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3903_cast_fp16 = reduce_mean(axes = var_3903_axes_0, keep_dims = var_3903_keep_dims_0, x = var_3901_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3905_cast_fp16 = add(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            fp32 var_3906_epsilon_0 = const()[name = string("op_3906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3906_cast_fp16 = rsqrt(epsilon = var_3906_epsilon_0, x = var_3905_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3907_cast_fp16 = mul(x = x_97_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437320448)))];
            tensor<fp16, [1, 16, 1024]> input_81_cast_fp16 = mul(x = var_3907_cast_fp16, y = model_layers_8_input_layernorm_weight_to_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437322560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439419776))))[name = string("model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_3916 = const()[name = string("op_3916"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3917_cast_fp16 = reshape(shape = var_3916, x = linear_56_cast_fp16)[name = string("op_3917_cast_fp16")];
            tensor<int32, [4]> x_99_perm_0 = const()[name = string("x_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3925_promoted_to_fp16 = const()[name = string("op_3925_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_99_cast_fp16 = transpose(perm = x_99_perm_0, x = var_3917_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_3927_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3925_promoted_to_fp16)[name = string("op_3927_cast_fp16")];
            tensor<int32, [1]> var_3929_axes_0 = const()[name = string("op_3929_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3929_keep_dims_0 = const()[name = string("op_3929_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3929_cast_fp16 = reduce_mean(axes = var_3929_axes_0, keep_dims = var_3929_keep_dims_0, x = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            fp16 var_3930_to_fp16 = const()[name = string("op_3930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3931_cast_fp16 = add(x = var_3929_cast_fp16, y = var_3930_to_fp16)[name = string("op_3931_cast_fp16")];
            fp32 var_3932_epsilon_0 = const()[name = string("op_3932_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3932_cast_fp16 = rsqrt(epsilon = var_3932_epsilon_0, x = var_3931_cast_fp16)[name = string("op_3932_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3933_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3932_cast_fp16)[name = string("op_3933_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439423936)))];
            tensor<fp16, [1, 16, 16, 128]> x_103_cast_fp16 = mul(x = var_3933_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_103_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439424256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440472896))))[name = string("model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3943_cast_fp16 = reshape(shape = var_3942, x = linear_57_cast_fp16)[name = string("op_3943_cast_fp16")];
            tensor<int32, [4]> x_101_perm_0 = const()[name = string("x_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3951_promoted_to_fp16 = const()[name = string("op_3951_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_101_cast_fp16 = transpose(perm = x_101_perm_0, x = var_3943_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_3953_cast_fp16 = pow(x = x_101_cast_fp16, y = var_3951_promoted_to_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]>([-1])];
            bool var_3955_keep_dims_0 = const()[name = string("op_3955_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3955_cast_fp16 = reduce_mean(axes = var_3955_axes_0, keep_dims = var_3955_keep_dims_0, x = var_3953_cast_fp16)[name = string("op_3955_cast_fp16")];
            fp16 var_3956_to_fp16 = const()[name = string("op_3956_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3957_cast_fp16 = add(x = var_3955_cast_fp16, y = var_3956_to_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3958_cast_fp16 = rsqrt(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3959_cast_fp16 = mul(x = x_101_cast_fp16, y = var_3958_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475008)))];
            tensor<fp16, [1, 8, 16, 128]> x_105_cast_fp16 = mul(x = var_3959_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441523968))))[name = string("model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_3968 = const()[name = string("op_3968"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3969_cast_fp16 = reshape(shape = var_3968, x = linear_58_cast_fp16)[name = string("op_3969_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_33_begin_0 = const()[name = string("first_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_33_end_0 = const()[name = string("first_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_33_end_mask_0 = const()[name = string("first_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_33_cast_fp16 = slice_by_index(begin = first_33_begin_0, end = first_33_end_0, end_mask = first_33_end_mask_0, x = x_103_cast_fp16)[name = string("first_33_cast_fp16")];
            tensor<int32, [4]> second_33_begin_0 = const()[name = string("second_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_33_end_0 = const()[name = string("second_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_33_end_mask_0 = const()[name = string("second_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_33_cast_fp16 = slice_by_index(begin = second_33_begin_0, end = second_33_end_0, end_mask = second_33_end_mask_0, x = x_103_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3983_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3984_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3986_cast_fp16 = sub(x = var_3983_cast_fp16, y = var_3984_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3987_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3988_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3990_cast_fp16 = add(x = var_3987_cast_fp16, y = var_3988_cast_fp16)[name = string("op_3990_cast_fp16")];
            int32 var_3992 = const()[name = string("op_3992"), val = int32(-1)];
            bool q_17_interleave_0 = const()[name = string("q_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_17_cast_fp16 = concat(axis = var_3992, interleave = q_17_interleave_0, values = (var_3986_cast_fp16, var_3990_cast_fp16))[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> first_35_begin_0 = const()[name = string("first_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_35_end_0 = const()[name = string("first_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_35_end_mask_0 = const()[name = string("first_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_35_cast_fp16 = slice_by_index(begin = first_35_begin_0, end = first_35_end_0, end_mask = first_35_end_mask_0, x = x_105_cast_fp16)[name = string("first_35_cast_fp16")];
            tensor<int32, [4]> second_35_begin_0 = const()[name = string("second_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_35_end_0 = const()[name = string("second_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_35_end_mask_0 = const()[name = string("second_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_35_cast_fp16 = slice_by_index(begin = second_35_begin_0, end = second_35_end_0, end_mask = second_35_end_mask_0, x = x_105_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4004_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4004_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4005_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4005_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4007_cast_fp16 = sub(x = var_4004_cast_fp16, y = var_4005_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4008_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4009_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4011_cast_fp16 = add(x = var_4008_cast_fp16, y = var_4009_cast_fp16)[name = string("op_4011_cast_fp16")];
            int32 var_4013 = const()[name = string("op_4013"), val = int32(-1)];
            bool k_17_interleave_0 = const()[name = string("k_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_17_cast_fp16 = concat(axis = var_4013, interleave = k_17_interleave_0, values = (var_4007_cast_fp16, var_4011_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4024_begin_0 = const()[name = string("op_4024_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4024_end_0 = const()[name = string("op_4024_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4024_stride_0 = const()[name = string("op_4024_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4024_cast_fp16 = slice_by_index(begin = var_4024_begin_0, end = var_4024_end_0, end_mask = var_4024_end_mask_0, stride = var_4024_stride_0, x = q_17_cast_fp16)[name = string("op_4024_cast_fp16")];
            tensor<int32, [4]> var_4034_begin_0 = const()[name = string("op_4034_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4034_end_0 = const()[name = string("op_4034_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4034_end_mask_0 = const()[name = string("op_4034_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4034_stride_0 = const()[name = string("op_4034_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4034_cast_fp16 = slice_by_index(begin = var_4034_begin_0, end = var_4034_end_0, end_mask = var_4034_end_mask_0, stride = var_4034_stride_0, x = q_17_cast_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4036 = const()[name = string("op_4036"), val = int32(2)];
            bool stacked_17_interleave_0 = const()[name = string("stacked_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_17_cast_fp16 = concat(axis = var_4036, interleave = stacked_17_interleave_0, values = (var_4024_cast_fp16, var_4034_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4042_begin_0 = const()[name = string("op_4042_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4042_end_0 = const()[name = string("op_4042_end_0"), val = tensor<int32, [4]>([9, 8, 4112, 128])];
            tensor<bool, [4]> var_4042_end_mask_0 = const()[name = string("op_4042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4042_cast_fp16 = slice_by_index(begin = var_4042_begin_0, end = var_4042_end_0, end_mask = var_4042_end_mask_0, x = read_state_0)[name = string("op_4042_cast_fp16")];
            tensor<int32, [4]> var_4052_begin_0 = const()[name = string("op_4052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4052_end_0 = const()[name = string("op_4052_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4052_end_mask_0 = const()[name = string("op_4052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4052_cast_fp16 = slice_by_index(begin = var_4052_begin_0, end = var_4052_end_0, end_mask = var_4052_end_mask_0, x = var_4042_cast_fp16)[name = string("op_4052_cast_fp16")];
            bool var_4061_transpose_x_1 = const()[name = string("op_4061_transpose_x_1"), val = bool(false)];
            bool var_4061_transpose_y_1 = const()[name = string("op_4061_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4061_cast_fp16 = matmul(transpose_x = var_4061_transpose_x_1, transpose_y = var_4061_transpose_y_1, x = var_4052_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4061_cast_fp16")];
            tensor<int32, [4]> var_4064_perm_0 = const()[name = string("op_4064_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4065_to_fp16 = const()[name = string("op_4065_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4064_cast_fp16 = transpose(perm = var_4064_perm_0, x = var_4061_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 4096]> var_4066_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_17_cast_fp16 = add(x = var_4066_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4072_transpose_x_1 = const()[name = string("op_4072_transpose_x_1"), val = bool(false)];
            bool var_4072_transpose_y_1 = const()[name = string("op_4072_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4072_cast_fp16 = matmul(transpose_x = var_4072_transpose_x_1, transpose_y = var_4072_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4072_cast_fp16")];
            fp16 var_4073_to_fp16 = const()[name = string("op_4073_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4074_cast_fp16 = mul(x = var_4072_cast_fp16, y = var_4073_to_fp16)[name = string("op_4074_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4074_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            bool var_4079_interleave_0 = const()[name = string("op_4079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4079_cast_fp16 = concat(axis = var_4078, interleave = var_4079_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4079_cast_fp16")];
            int32 var_4080 = const()[name = string("op_4080"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_17_cast_fp16 = softmax(axis = var_4080, x = var_4079_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4087_begin_0 = const()[name = string("op_4087_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4087_end_0 = const()[name = string("op_4087_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4087_end_mask_0 = const()[name = string("op_4087_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4087_cast_fp16 = slice_by_index(begin = var_4087_begin_0, end = var_4087_end_0, end_mask = var_4087_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4087_cast_fp16")];
            tensor<int32, [4]> var_4092_begin_0 = const()[name = string("op_4092_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4092_end_0 = const()[name = string("op_4092_end_0"), val = tensor<int32, [4]>([9, 8, 4112, 128])];
            tensor<bool, [4]> var_4092_end_mask_0 = const()[name = string("op_4092_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4092_cast_fp16 = slice_by_index(begin = var_4092_begin_0, end = var_4092_end_0, end_mask = var_4092_end_mask_0, x = read_state_1)[name = string("op_4092_cast_fp16")];
            tensor<int32, [4]> var_4102_begin_0 = const()[name = string("op_4102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4102_end_0 = const()[name = string("op_4102_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4102_end_mask_0 = const()[name = string("op_4102_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4102_cast_fp16 = slice_by_index(begin = var_4102_begin_0, end = var_4102_end_0, end_mask = var_4102_end_mask_0, x = var_4092_cast_fp16)[name = string("op_4102_cast_fp16")];
            bool var_4108_transpose_x_0 = const()[name = string("op_4108_transpose_x_0"), val = bool(false)];
            bool var_4108_transpose_y_0 = const()[name = string("op_4108_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4108_cast_fp16 = matmul(transpose_x = var_4108_transpose_x_0, transpose_y = var_4108_transpose_y_0, x = var_4087_cast_fp16, y = var_4102_cast_fp16)[name = string("op_4108_cast_fp16")];
            tensor<int32, [4]> var_4113_begin_0 = const()[name = string("op_4113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4113_end_0 = const()[name = string("op_4113_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4113_end_mask_0 = const()[name = string("op_4113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4113_cast_fp16 = slice_by_index(begin = var_4113_begin_0, end = var_4113_end_0, end_mask = var_4113_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4113_cast_fp16")];
            bool var_4114_transpose_x_0 = const()[name = string("op_4114_transpose_x_0"), val = bool(false)];
            bool var_4114_transpose_y_0 = const()[name = string("op_4114_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_3969_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4114_cast_fp16 = matmul(transpose_x = var_4114_transpose_x_0, transpose_y = var_4114_transpose_y_0, x = var_4113_cast_fp16, y = v_17_cast_fp16)[name = string("op_4114_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4108_cast_fp16, y = var_4114_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4131_begin_0 = const()[name = string("op_4131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4131_end_0 = const()[name = string("op_4131_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4131_end_mask_0 = const()[name = string("op_4131_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4131_cast_fp16 = slice_by_index(begin = var_4131_begin_0, end = var_4131_end_0, end_mask = var_4131_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<int32, [4]> var_4146_begin_0 = const()[name = string("op_4146_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4146_end_0 = const()[name = string("op_4146_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4146_end_mask_0 = const()[name = string("op_4146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4146_cast_fp16 = slice_by_index(begin = var_4146_begin_0, end = var_4146_end_0, end_mask = var_4146_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4146_cast_fp16")];
            int32 var_4148 = const()[name = string("op_4148"), val = int32(1)];
            bool mixed_35_interleave_0 = const()[name = string("mixed_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_35_cast_fp16 = concat(axis = var_4148, interleave = mixed_35_interleave_0, values = (var_4131_cast_fp16, var_4146_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4152_perm_0 = const()[name = string("op_4152_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4156 = const()[name = string("op_4156"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4152_cast_fp16 = transpose(perm = var_4152_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4156, x = var_4152_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441526080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443623296))))[name = string("model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_107_cast_fp16 = add(x = x_97_cast_fp16, y = linear_59_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_4167_promoted_to_fp16 = const()[name = string("op_4167_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4169_cast_fp16 = pow(x = x_107_cast_fp16, y = var_4167_promoted_to_fp16)[name = string("op_4169_cast_fp16")];
            tensor<int32, [1]> var_4171_axes_0 = const()[name = string("op_4171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4171_keep_dims_0 = const()[name = string("op_4171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4171_cast_fp16 = reduce_mean(axes = var_4171_axes_0, keep_dims = var_4171_keep_dims_0, x = var_4169_cast_fp16)[name = string("op_4171_cast_fp16")];
            fp16 var_4172_to_fp16 = const()[name = string("op_4172_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4173_cast_fp16 = add(x = var_4171_cast_fp16, y = var_4172_to_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4174_cast_fp16 = rsqrt(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4175_cast_fp16 = mul(x = x_107_cast_fp16, y = var_4174_cast_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443625408)))];
            tensor<fp16, [1, 16, 1024]> input_85_cast_fp16 = mul(x = var_4175_cast_fp16, y = model_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443627520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446773312))))[name = string("model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4180_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446779520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449925312))))[name = string("model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_89_cast_fp16 = mul(x = var_4180_cast_fp16, y = linear_61_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449931520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453077312))))[name = string("model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = linear_62_cast_fp16)[name = string("x_109_cast_fp16")];
            fp16 var_4194_promoted_to_fp16 = const()[name = string("op_4194_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4196_cast_fp16 = pow(x = x_109_cast_fp16, y = var_4194_promoted_to_fp16)[name = string("op_4196_cast_fp16")];
            tensor<int32, [1]> var_4198_axes_0 = const()[name = string("op_4198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4198_keep_dims_0 = const()[name = string("op_4198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4198_cast_fp16 = reduce_mean(axes = var_4198_axes_0, keep_dims = var_4198_keep_dims_0, x = var_4196_cast_fp16)[name = string("op_4198_cast_fp16")];
            fp16 var_4199_to_fp16 = const()[name = string("op_4199_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4200_cast_fp16 = add(x = var_4198_cast_fp16, y = var_4199_to_fp16)[name = string("op_4200_cast_fp16")];
            fp32 var_4201_epsilon_0 = const()[name = string("op_4201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4201_cast_fp16 = rsqrt(epsilon = var_4201_epsilon_0, x = var_4200_cast_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4202_cast_fp16 = mul(x = x_109_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453079424)))];
            tensor<fp16, [1, 16, 1024]> input_91_cast_fp16 = mul(x = var_4202_cast_fp16, y = model_layers_9_input_layernorm_weight_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453081536))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455178752))))[name = string("model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_4211 = const()[name = string("op_4211"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4212_cast_fp16 = reshape(shape = var_4211, x = linear_63_cast_fp16)[name = string("op_4212_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4220_promoted_to_fp16 = const()[name = string("op_4220_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4212_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4222_cast_fp16 = pow(x = x_111_cast_fp16, y = var_4220_promoted_to_fp16)[name = string("op_4222_cast_fp16")];
            tensor<int32, [1]> var_4224_axes_0 = const()[name = string("op_4224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4224_keep_dims_0 = const()[name = string("op_4224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4224_cast_fp16 = reduce_mean(axes = var_4224_axes_0, keep_dims = var_4224_keep_dims_0, x = var_4222_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp16 var_4225_to_fp16 = const()[name = string("op_4225_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4226_cast_fp16 = add(x = var_4224_cast_fp16, y = var_4225_to_fp16)[name = string("op_4226_cast_fp16")];
            fp32 var_4227_epsilon_0 = const()[name = string("op_4227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4227_cast_fp16 = rsqrt(epsilon = var_4227_epsilon_0, x = var_4226_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4228_cast_fp16 = mul(x = x_111_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455182912)))];
            tensor<fp16, [1, 16, 16, 128]> x_115_cast_fp16 = mul(x = var_4228_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455183232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456231872))))[name = string("model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4237 = const()[name = string("op_4237"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4238_cast_fp16 = reshape(shape = var_4237, x = linear_64_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4246_promoted_to_fp16 = const()[name = string("op_4246_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4238_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4248_cast_fp16 = pow(x = x_113_cast_fp16, y = var_4246_promoted_to_fp16)[name = string("op_4248_cast_fp16")];
            tensor<int32, [1]> var_4250_axes_0 = const()[name = string("op_4250_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4250_keep_dims_0 = const()[name = string("op_4250_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4250_cast_fp16 = reduce_mean(axes = var_4250_axes_0, keep_dims = var_4250_keep_dims_0, x = var_4248_cast_fp16)[name = string("op_4250_cast_fp16")];
            fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4252_cast_fp16 = add(x = var_4250_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")];
            fp32 var_4253_epsilon_0 = const()[name = string("op_4253_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4253_cast_fp16 = rsqrt(epsilon = var_4253_epsilon_0, x = var_4252_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4254_cast_fp16 = mul(x = x_113_cast_fp16, y = var_4253_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456233984)))];
            tensor<fp16, [1, 8, 16, 128]> x_117_cast_fp16 = mul(x = var_4254_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456234304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457282944))))[name = string("model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4263 = const()[name = string("op_4263"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4264_cast_fp16 = reshape(shape = var_4263, x = linear_65_cast_fp16)[name = string("op_4264_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_37_begin_0 = const()[name = string("first_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_37_end_0 = const()[name = string("first_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_37_end_mask_0 = const()[name = string("first_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_37_cast_fp16 = slice_by_index(begin = first_37_begin_0, end = first_37_end_0, end_mask = first_37_end_mask_0, x = x_115_cast_fp16)[name = string("first_37_cast_fp16")];
            tensor<int32, [4]> second_37_begin_0 = const()[name = string("second_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_37_end_0 = const()[name = string("second_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_37_end_mask_0 = const()[name = string("second_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_37_cast_fp16 = slice_by_index(begin = second_37_begin_0, end = second_37_end_0, end_mask = second_37_end_mask_0, x = x_115_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4278_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4278_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4279_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4279_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4281_cast_fp16 = sub(x = var_4278_cast_fp16, y = var_4279_cast_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4282_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4283_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4283_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4285_cast_fp16 = add(x = var_4282_cast_fp16, y = var_4283_cast_fp16)[name = string("op_4285_cast_fp16")];
            int32 var_4287 = const()[name = string("op_4287"), val = int32(-1)];
            bool q_19_interleave_0 = const()[name = string("q_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_19_cast_fp16 = concat(axis = var_4287, interleave = q_19_interleave_0, values = (var_4281_cast_fp16, var_4285_cast_fp16))[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> first_39_begin_0 = const()[name = string("first_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_39_end_0 = const()[name = string("first_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_39_end_mask_0 = const()[name = string("first_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_39_cast_fp16 = slice_by_index(begin = first_39_begin_0, end = first_39_end_0, end_mask = first_39_end_mask_0, x = x_117_cast_fp16)[name = string("first_39_cast_fp16")];
            tensor<int32, [4]> second_39_begin_0 = const()[name = string("second_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_39_end_0 = const()[name = string("second_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_39_end_mask_0 = const()[name = string("second_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_39_cast_fp16 = slice_by_index(begin = second_39_begin_0, end = second_39_end_0, end_mask = second_39_end_mask_0, x = x_117_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4299_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4299_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4300_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4302_cast_fp16 = sub(x = var_4299_cast_fp16, y = var_4300_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4303_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4303_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4304_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4304_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4306_cast_fp16 = add(x = var_4303_cast_fp16, y = var_4304_cast_fp16)[name = string("op_4306_cast_fp16")];
            int32 var_4308 = const()[name = string("op_4308"), val = int32(-1)];
            bool k_19_interleave_0 = const()[name = string("k_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_19_cast_fp16 = concat(axis = var_4308, interleave = k_19_interleave_0, values = (var_4302_cast_fp16, var_4306_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4319_begin_0 = const()[name = string("op_4319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4319_end_0 = const()[name = string("op_4319_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4319_end_mask_0 = const()[name = string("op_4319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4319_stride_0 = const()[name = string("op_4319_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4319_cast_fp16 = slice_by_index(begin = var_4319_begin_0, end = var_4319_end_0, end_mask = var_4319_end_mask_0, stride = var_4319_stride_0, x = q_19_cast_fp16)[name = string("op_4319_cast_fp16")];
            tensor<int32, [4]> var_4329_begin_0 = const()[name = string("op_4329_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4329_end_0 = const()[name = string("op_4329_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4329_end_mask_0 = const()[name = string("op_4329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4329_stride_0 = const()[name = string("op_4329_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4329_cast_fp16 = slice_by_index(begin = var_4329_begin_0, end = var_4329_end_0, end_mask = var_4329_end_mask_0, stride = var_4329_stride_0, x = q_19_cast_fp16)[name = string("op_4329_cast_fp16")];
            int32 var_4331 = const()[name = string("op_4331"), val = int32(2)];
            bool stacked_19_interleave_0 = const()[name = string("stacked_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_19_cast_fp16 = concat(axis = var_4331, interleave = stacked_19_interleave_0, values = (var_4319_cast_fp16, var_4329_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([10, 8, 4112, 128])];
            tensor<bool, [4]> var_4337_end_mask_0 = const()[name = string("op_4337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4337_cast_fp16 = slice_by_index(begin = var_4337_begin_0, end = var_4337_end_0, end_mask = var_4337_end_mask_0, x = read_state_0)[name = string("op_4337_cast_fp16")];
            tensor<int32, [4]> var_4347_begin_0 = const()[name = string("op_4347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4347_end_0 = const()[name = string("op_4347_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4347_end_mask_0 = const()[name = string("op_4347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4347_cast_fp16 = slice_by_index(begin = var_4347_begin_0, end = var_4347_end_0, end_mask = var_4347_end_mask_0, x = var_4337_cast_fp16)[name = string("op_4347_cast_fp16")];
            bool var_4356_transpose_x_1 = const()[name = string("op_4356_transpose_x_1"), val = bool(false)];
            bool var_4356_transpose_y_1 = const()[name = string("op_4356_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4356_cast_fp16 = matmul(transpose_x = var_4356_transpose_x_1, transpose_y = var_4356_transpose_y_1, x = var_4347_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4356_cast_fp16")];
            tensor<int32, [4]> var_4359_perm_0 = const()[name = string("op_4359_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4360_to_fp16 = const()[name = string("op_4360_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4359_cast_fp16 = transpose(perm = var_4359_perm_0, x = var_4356_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 4096]> var_4361_cast_fp16 = mul(x = var_4359_cast_fp16, y = var_4360_to_fp16)[name = string("op_4361_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_19_cast_fp16 = add(x = var_4361_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4367_transpose_x_1 = const()[name = string("op_4367_transpose_x_1"), val = bool(false)];
            bool var_4367_transpose_y_1 = const()[name = string("op_4367_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4367_cast_fp16 = matmul(transpose_x = var_4367_transpose_x_1, transpose_y = var_4367_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4367_cast_fp16")];
            fp16 var_4368_to_fp16 = const()[name = string("op_4368_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4369_cast_fp16 = mul(x = var_4367_cast_fp16, y = var_4368_to_fp16)[name = string("op_4369_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4369_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4373 = const()[name = string("op_4373"), val = int32(-1)];
            bool var_4374_interleave_0 = const()[name = string("op_4374_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4374_cast_fp16 = concat(axis = var_4373, interleave = var_4374_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4374_cast_fp16")];
            int32 var_4375 = const()[name = string("op_4375"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_19_cast_fp16 = softmax(axis = var_4375, x = var_4374_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4382_begin_0 = const()[name = string("op_4382_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4382_end_0 = const()[name = string("op_4382_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4382_end_mask_0 = const()[name = string("op_4382_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4382_cast_fp16 = slice_by_index(begin = var_4382_begin_0, end = var_4382_end_0, end_mask = var_4382_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4382_cast_fp16")];
            tensor<int32, [4]> var_4387_begin_0 = const()[name = string("op_4387_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4387_end_0 = const()[name = string("op_4387_end_0"), val = tensor<int32, [4]>([10, 8, 4112, 128])];
            tensor<bool, [4]> var_4387_end_mask_0 = const()[name = string("op_4387_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4387_cast_fp16 = slice_by_index(begin = var_4387_begin_0, end = var_4387_end_0, end_mask = var_4387_end_mask_0, x = read_state_1)[name = string("op_4387_cast_fp16")];
            tensor<int32, [4]> var_4397_begin_0 = const()[name = string("op_4397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4397_end_0 = const()[name = string("op_4397_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4397_end_mask_0 = const()[name = string("op_4397_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4397_cast_fp16 = slice_by_index(begin = var_4397_begin_0, end = var_4397_end_0, end_mask = var_4397_end_mask_0, x = var_4387_cast_fp16)[name = string("op_4397_cast_fp16")];
            bool var_4403_transpose_x_0 = const()[name = string("op_4403_transpose_x_0"), val = bool(false)];
            bool var_4403_transpose_y_0 = const()[name = string("op_4403_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4403_cast_fp16 = matmul(transpose_x = var_4403_transpose_x_0, transpose_y = var_4403_transpose_y_0, x = var_4382_cast_fp16, y = var_4397_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<int32, [4]> var_4408_begin_0 = const()[name = string("op_4408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4408_end_0 = const()[name = string("op_4408_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4408_end_mask_0 = const()[name = string("op_4408_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4408_cast_fp16")];
            bool var_4409_transpose_x_0 = const()[name = string("op_4409_transpose_x_0"), val = bool(false)];
            bool var_4409_transpose_y_0 = const()[name = string("op_4409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4264_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4409_cast_fp16 = matmul(transpose_x = var_4409_transpose_x_0, transpose_y = var_4409_transpose_y_0, x = var_4408_cast_fp16, y = v_19_cast_fp16)[name = string("op_4409_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4403_cast_fp16, y = var_4409_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4426_begin_0 = const()[name = string("op_4426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4426_end_0 = const()[name = string("op_4426_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4426_end_mask_0 = const()[name = string("op_4426_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4426_cast_fp16 = slice_by_index(begin = var_4426_begin_0, end = var_4426_end_0, end_mask = var_4426_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<int32, [4]> var_4441_begin_0 = const()[name = string("op_4441_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4441_end_0 = const()[name = string("op_4441_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4441_end_mask_0 = const()[name = string("op_4441_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4441_cast_fp16 = slice_by_index(begin = var_4441_begin_0, end = var_4441_end_0, end_mask = var_4441_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4441_cast_fp16")];
            int32 var_4443 = const()[name = string("op_4443"), val = int32(1)];
            bool mixed_39_interleave_0 = const()[name = string("mixed_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_39_cast_fp16 = concat(axis = var_4443, interleave = mixed_39_interleave_0, values = (var_4426_cast_fp16, var_4441_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4447_perm_0 = const()[name = string("op_4447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4451 = const()[name = string("op_4451"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4447_cast_fp16 = transpose(perm = var_4447_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4451, x = var_4447_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457285056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459382272))))[name = string("model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = linear_66_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_4462_promoted_to_fp16 = const()[name = string("op_4462_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4464_cast_fp16 = pow(x = x_119_cast_fp16, y = var_4462_promoted_to_fp16)[name = string("op_4464_cast_fp16")];
            tensor<int32, [1]> var_4466_axes_0 = const()[name = string("op_4466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4466_keep_dims_0 = const()[name = string("op_4466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4466_cast_fp16 = reduce_mean(axes = var_4466_axes_0, keep_dims = var_4466_keep_dims_0, x = var_4464_cast_fp16)[name = string("op_4466_cast_fp16")];
            fp16 var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4468_cast_fp16 = add(x = var_4466_cast_fp16, y = var_4467_to_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4470_cast_fp16 = mul(x = x_119_cast_fp16, y = var_4469_cast_fp16)[name = string("op_4470_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459384384)))];
            tensor<fp16, [1, 16, 1024]> input_95_cast_fp16 = mul(x = var_4470_cast_fp16, y = model_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459386496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462532288))))[name = string("model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4475_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462538496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465684288))))[name = string("model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_99_cast_fp16 = mul(x = var_4475_cast_fp16, y = linear_68_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465690496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468836288))))[name = string("model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_121_cast_fp16 = add(x = x_119_cast_fp16, y = linear_69_cast_fp16)[name = string("x_121_cast_fp16")];
            fp16 var_4489_promoted_to_fp16 = const()[name = string("op_4489_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4491_cast_fp16 = pow(x = x_121_cast_fp16, y = var_4489_promoted_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4493_keep_dims_0 = const()[name = string("op_4493_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4493_cast_fp16 = reduce_mean(axes = var_4493_axes_0, keep_dims = var_4493_keep_dims_0, x = var_4491_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4495_cast_fp16 = add(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("op_4495_cast_fp16")];
            fp32 var_4496_epsilon_0 = const()[name = string("op_4496_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4496_cast_fp16 = rsqrt(epsilon = var_4496_epsilon_0, x = var_4495_cast_fp16)[name = string("op_4496_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4497_cast_fp16 = mul(x = x_121_cast_fp16, y = var_4496_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468838400)))];
            tensor<fp16, [1, 16, 1024]> input_101_cast_fp16 = mul(x = var_4497_cast_fp16, y = model_layers_10_input_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468840512))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470937728))))[name = string("model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4507_cast_fp16 = reshape(shape = var_4506, x = linear_70_cast_fp16)[name = string("op_4507_cast_fp16")];
            tensor<int32, [4]> x_123_perm_0 = const()[name = string("x_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4515_promoted_to_fp16 = const()[name = string("op_4515_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_123_cast_fp16 = transpose(perm = x_123_perm_0, x = var_4507_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4517_cast_fp16 = pow(x = x_123_cast_fp16, y = var_4515_promoted_to_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4519_axes_0 = const()[name = string("op_4519_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4519_keep_dims_0 = const()[name = string("op_4519_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4519_cast_fp16 = reduce_mean(axes = var_4519_axes_0, keep_dims = var_4519_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4519_cast_fp16")];
            fp16 var_4520_to_fp16 = const()[name = string("op_4520_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4521_cast_fp16 = add(x = var_4519_cast_fp16, y = var_4520_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4522_cast_fp16 = rsqrt(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4523_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4522_cast_fp16)[name = string("op_4523_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470941888)))];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = mul(x = var_4523_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470942208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471990848))))[name = string("model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4532 = const()[name = string("op_4532"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4533_cast_fp16 = reshape(shape = var_4532, x = linear_71_cast_fp16)[name = string("op_4533_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4541_promoted_to_fp16 = const()[name = string("op_4541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_4533_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4543_cast_fp16 = pow(x = x_125_cast_fp16, y = var_4541_promoted_to_fp16)[name = string("op_4543_cast_fp16")];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4545_keep_dims_0 = const()[name = string("op_4545_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4545_cast_fp16 = reduce_mean(axes = var_4545_axes_0, keep_dims = var_4545_keep_dims_0, x = var_4543_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4547_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("op_4547_cast_fp16")];
            fp32 var_4548_epsilon_0 = const()[name = string("op_4548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4548_cast_fp16 = rsqrt(epsilon = var_4548_epsilon_0, x = var_4547_cast_fp16)[name = string("op_4548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4549_cast_fp16 = mul(x = x_125_cast_fp16, y = var_4548_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471992960)))];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = mul(x = var_4549_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471993280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473041920))))[name = string("model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4558 = const()[name = string("op_4558"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4559_cast_fp16 = reshape(shape = var_4558, x = linear_72_cast_fp16)[name = string("op_4559_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_41_begin_0 = const()[name = string("first_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_41_end_0 = const()[name = string("first_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_41_end_mask_0 = const()[name = string("first_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_41_cast_fp16 = slice_by_index(begin = first_41_begin_0, end = first_41_end_0, end_mask = first_41_end_mask_0, x = x_127_cast_fp16)[name = string("first_41_cast_fp16")];
            tensor<int32, [4]> second_41_begin_0 = const()[name = string("second_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_41_end_0 = const()[name = string("second_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_41_end_mask_0 = const()[name = string("second_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_41_cast_fp16 = slice_by_index(begin = second_41_begin_0, end = second_41_end_0, end_mask = second_41_end_mask_0, x = x_127_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4573_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4574_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4574_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4576_cast_fp16 = sub(x = var_4573_cast_fp16, y = var_4574_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4577_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4578_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4578_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4580_cast_fp16 = add(x = var_4577_cast_fp16, y = var_4578_cast_fp16)[name = string("op_4580_cast_fp16")];
            int32 var_4582 = const()[name = string("op_4582"), val = int32(-1)];
            bool q_21_interleave_0 = const()[name = string("q_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_21_cast_fp16 = concat(axis = var_4582, interleave = q_21_interleave_0, values = (var_4576_cast_fp16, var_4580_cast_fp16))[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> first_43_begin_0 = const()[name = string("first_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_43_end_0 = const()[name = string("first_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_43_end_mask_0 = const()[name = string("first_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_43_cast_fp16 = slice_by_index(begin = first_43_begin_0, end = first_43_end_0, end_mask = first_43_end_mask_0, x = x_129_cast_fp16)[name = string("first_43_cast_fp16")];
            tensor<int32, [4]> second_43_begin_0 = const()[name = string("second_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_43_end_0 = const()[name = string("second_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_43_end_mask_0 = const()[name = string("second_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_43_cast_fp16 = slice_by_index(begin = second_43_begin_0, end = second_43_end_0, end_mask = second_43_end_mask_0, x = x_129_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4594_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4594_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4595_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4595_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4597_cast_fp16 = sub(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4598_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4599_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4599_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4601_cast_fp16 = add(x = var_4598_cast_fp16, y = var_4599_cast_fp16)[name = string("op_4601_cast_fp16")];
            int32 var_4603 = const()[name = string("op_4603"), val = int32(-1)];
            bool k_21_interleave_0 = const()[name = string("k_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_21_cast_fp16 = concat(axis = var_4603, interleave = k_21_interleave_0, values = (var_4597_cast_fp16, var_4601_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4614_begin_0 = const()[name = string("op_4614_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4614_end_0 = const()[name = string("op_4614_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4614_end_mask_0 = const()[name = string("op_4614_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4614_stride_0 = const()[name = string("op_4614_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4614_cast_fp16 = slice_by_index(begin = var_4614_begin_0, end = var_4614_end_0, end_mask = var_4614_end_mask_0, stride = var_4614_stride_0, x = q_21_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<int32, [4]> var_4624_begin_0 = const()[name = string("op_4624_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4624_end_0 = const()[name = string("op_4624_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4624_end_mask_0 = const()[name = string("op_4624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4624_stride_0 = const()[name = string("op_4624_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4624_cast_fp16 = slice_by_index(begin = var_4624_begin_0, end = var_4624_end_0, end_mask = var_4624_end_mask_0, stride = var_4624_stride_0, x = q_21_cast_fp16)[name = string("op_4624_cast_fp16")];
            int32 var_4626 = const()[name = string("op_4626"), val = int32(2)];
            bool stacked_21_interleave_0 = const()[name = string("stacked_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_21_cast_fp16 = concat(axis = var_4626, interleave = stacked_21_interleave_0, values = (var_4614_cast_fp16, var_4624_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4632_begin_0 = const()[name = string("op_4632_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4632_end_0 = const()[name = string("op_4632_end_0"), val = tensor<int32, [4]>([11, 8, 4112, 128])];
            tensor<bool, [4]> var_4632_end_mask_0 = const()[name = string("op_4632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4632_cast_fp16 = slice_by_index(begin = var_4632_begin_0, end = var_4632_end_0, end_mask = var_4632_end_mask_0, x = read_state_0)[name = string("op_4632_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = var_4632_cast_fp16)[name = string("op_4642_cast_fp16")];
            bool var_4651_transpose_x_1 = const()[name = string("op_4651_transpose_x_1"), val = bool(false)];
            bool var_4651_transpose_y_1 = const()[name = string("op_4651_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_1, transpose_y = var_4651_transpose_y_1, x = var_4642_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [4]> var_4654_perm_0 = const()[name = string("op_4654_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4655_to_fp16 = const()[name = string("op_4655_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4654_cast_fp16 = transpose(perm = var_4654_perm_0, x = var_4651_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 4096]> var_4656_cast_fp16 = mul(x = var_4654_cast_fp16, y = var_4655_to_fp16)[name = string("op_4656_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_21_cast_fp16 = add(x = var_4656_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4662_transpose_x_1 = const()[name = string("op_4662_transpose_x_1"), val = bool(false)];
            bool var_4662_transpose_y_1 = const()[name = string("op_4662_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4662_cast_fp16 = matmul(transpose_x = var_4662_transpose_x_1, transpose_y = var_4662_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4662_cast_fp16")];
            fp16 var_4663_to_fp16 = const()[name = string("op_4663_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4664_cast_fp16 = mul(x = var_4662_cast_fp16, y = var_4663_to_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4664_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4668 = const()[name = string("op_4668"), val = int32(-1)];
            bool var_4669_interleave_0 = const()[name = string("op_4669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4669_cast_fp16 = concat(axis = var_4668, interleave = var_4669_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4669_cast_fp16")];
            int32 var_4670 = const()[name = string("op_4670"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_21_cast_fp16 = softmax(axis = var_4670, x = var_4669_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4677_begin_0 = const()[name = string("op_4677_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4677_end_0 = const()[name = string("op_4677_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4677_end_mask_0 = const()[name = string("op_4677_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4677_cast_fp16 = slice_by_index(begin = var_4677_begin_0, end = var_4677_end_0, end_mask = var_4677_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4677_cast_fp16")];
            tensor<int32, [4]> var_4682_begin_0 = const()[name = string("op_4682_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4682_end_0 = const()[name = string("op_4682_end_0"), val = tensor<int32, [4]>([11, 8, 4112, 128])];
            tensor<bool, [4]> var_4682_end_mask_0 = const()[name = string("op_4682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4682_cast_fp16 = slice_by_index(begin = var_4682_begin_0, end = var_4682_end_0, end_mask = var_4682_end_mask_0, x = read_state_1)[name = string("op_4682_cast_fp16")];
            tensor<int32, [4]> var_4692_begin_0 = const()[name = string("op_4692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4692_end_0 = const()[name = string("op_4692_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4692_end_mask_0 = const()[name = string("op_4692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4692_cast_fp16 = slice_by_index(begin = var_4692_begin_0, end = var_4692_end_0, end_mask = var_4692_end_mask_0, x = var_4682_cast_fp16)[name = string("op_4692_cast_fp16")];
            bool var_4698_transpose_x_0 = const()[name = string("op_4698_transpose_x_0"), val = bool(false)];
            bool var_4698_transpose_y_0 = const()[name = string("op_4698_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4698_cast_fp16 = matmul(transpose_x = var_4698_transpose_x_0, transpose_y = var_4698_transpose_y_0, x = var_4677_cast_fp16, y = var_4692_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<int32, [4]> var_4703_begin_0 = const()[name = string("op_4703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4703_end_0 = const()[name = string("op_4703_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4703_end_mask_0 = const()[name = string("op_4703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4703_cast_fp16 = slice_by_index(begin = var_4703_begin_0, end = var_4703_end_0, end_mask = var_4703_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4703_cast_fp16")];
            bool var_4704_transpose_x_0 = const()[name = string("op_4704_transpose_x_0"), val = bool(false)];
            bool var_4704_transpose_y_0 = const()[name = string("op_4704_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4559_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_4704_cast_fp16 = matmul(transpose_x = var_4704_transpose_x_0, transpose_y = var_4704_transpose_y_0, x = var_4703_cast_fp16, y = v_21_cast_fp16)[name = string("op_4704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4704_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_4721_begin_0 = const()[name = string("op_4721_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4721_end_0 = const()[name = string("op_4721_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4721_end_mask_0 = const()[name = string("op_4721_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4721_cast_fp16 = slice_by_index(begin = var_4721_begin_0, end = var_4721_end_0, end_mask = var_4721_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<int32, [4]> var_4736_begin_0 = const()[name = string("op_4736_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4736_end_0 = const()[name = string("op_4736_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4736_end_mask_0 = const()[name = string("op_4736_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4736_cast_fp16 = slice_by_index(begin = var_4736_begin_0, end = var_4736_end_0, end_mask = var_4736_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4736_cast_fp16")];
            int32 var_4738 = const()[name = string("op_4738"), val = int32(1)];
            bool mixed_43_interleave_0 = const()[name = string("mixed_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_43_cast_fp16 = concat(axis = var_4738, interleave = mixed_43_interleave_0, values = (var_4721_cast_fp16, var_4736_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_4742_perm_0 = const()[name = string("op_4742_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4746 = const()[name = string("op_4746"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4742_cast_fp16 = transpose(perm = var_4742_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_4746, x = var_4742_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473044032))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475141248))))[name = string("model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_131_cast_fp16 = add(x = x_121_cast_fp16, y = linear_73_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_4757_promoted_to_fp16 = const()[name = string("op_4757_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4759_cast_fp16 = pow(x = x_131_cast_fp16, y = var_4757_promoted_to_fp16)[name = string("op_4759_cast_fp16")];
            tensor<int32, [1]> var_4761_axes_0 = const()[name = string("op_4761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4761_keep_dims_0 = const()[name = string("op_4761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4761_cast_fp16 = reduce_mean(axes = var_4761_axes_0, keep_dims = var_4761_keep_dims_0, x = var_4759_cast_fp16)[name = string("op_4761_cast_fp16")];
            fp16 var_4762_to_fp16 = const()[name = string("op_4762_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4763_cast_fp16 = add(x = var_4761_cast_fp16, y = var_4762_to_fp16)[name = string("op_4763_cast_fp16")];
            fp32 var_4764_epsilon_0 = const()[name = string("op_4764_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4764_cast_fp16 = rsqrt(epsilon = var_4764_epsilon_0, x = var_4763_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_4764_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475143360)))];
            tensor<fp16, [1, 16, 1024]> input_105_cast_fp16 = mul(x = var_4765_cast_fp16, y = model_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475145472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478291264))))[name = string("model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4770_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_4770_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478297472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481443264))))[name = string("model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_109_cast_fp16 = mul(x = var_4770_cast_fp16, y = linear_75_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481449472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484595264))))[name = string("model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_133_cast_fp16 = add(x = x_131_cast_fp16, y = linear_76_cast_fp16)[name = string("x_133_cast_fp16")];
            fp16 var_4784_promoted_to_fp16 = const()[name = string("op_4784_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4786_cast_fp16 = pow(x = x_133_cast_fp16, y = var_4784_promoted_to_fp16)[name = string("op_4786_cast_fp16")];
            tensor<int32, [1]> var_4788_axes_0 = const()[name = string("op_4788_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4788_keep_dims_0 = const()[name = string("op_4788_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4788_cast_fp16 = reduce_mean(axes = var_4788_axes_0, keep_dims = var_4788_keep_dims_0, x = var_4786_cast_fp16)[name = string("op_4788_cast_fp16")];
            fp16 var_4789_to_fp16 = const()[name = string("op_4789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4790_cast_fp16 = add(x = var_4788_cast_fp16, y = var_4789_to_fp16)[name = string("op_4790_cast_fp16")];
            fp32 var_4791_epsilon_0 = const()[name = string("op_4791_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4791_cast_fp16 = rsqrt(epsilon = var_4791_epsilon_0, x = var_4790_cast_fp16)[name = string("op_4791_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4792_cast_fp16 = mul(x = x_133_cast_fp16, y = var_4791_cast_fp16)[name = string("op_4792_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484597376)))];
            tensor<fp16, [1, 16, 1024]> input_111_cast_fp16 = mul(x = var_4792_cast_fp16, y = model_layers_11_input_layernorm_weight_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484599488))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486696704))))[name = string("model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_4801 = const()[name = string("op_4801"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4802_cast_fp16 = reshape(shape = var_4801, x = linear_77_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4810_promoted_to_fp16 = const()[name = string("op_4810_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_4802_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_4812_cast_fp16 = pow(x = x_135_cast_fp16, y = var_4810_promoted_to_fp16)[name = string("op_4812_cast_fp16")];
            tensor<int32, [1]> var_4814_axes_0 = const()[name = string("op_4814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4814_keep_dims_0 = const()[name = string("op_4814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = reduce_mean(axes = var_4814_axes_0, keep_dims = var_4814_keep_dims_0, x = var_4812_cast_fp16)[name = string("op_4814_cast_fp16")];
            fp16 var_4815_to_fp16 = const()[name = string("op_4815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = add(x = var_4814_cast_fp16, y = var_4815_to_fp16)[name = string("op_4816_cast_fp16")];
            fp32 var_4817_epsilon_0 = const()[name = string("op_4817_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4817_cast_fp16 = rsqrt(epsilon = var_4817_epsilon_0, x = var_4816_cast_fp16)[name = string("op_4817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4818_cast_fp16 = mul(x = x_135_cast_fp16, y = var_4817_cast_fp16)[name = string("op_4818_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486700864)))];
            tensor<fp16, [1, 16, 16, 128]> x_139_cast_fp16 = mul(x = var_4818_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486701184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487749824))))[name = string("model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_4827 = const()[name = string("op_4827"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4828_cast_fp16 = reshape(shape = var_4827, x = linear_78_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<int32, [4]> x_137_perm_0 = const()[name = string("x_137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4836_promoted_to_fp16 = const()[name = string("op_4836_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_137_cast_fp16 = transpose(perm = x_137_perm_0, x = var_4828_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_4838_cast_fp16 = pow(x = x_137_cast_fp16, y = var_4836_promoted_to_fp16)[name = string("op_4838_cast_fp16")];
            tensor<int32, [1]> var_4840_axes_0 = const()[name = string("op_4840_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4840_keep_dims_0 = const()[name = string("op_4840_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4840_cast_fp16 = reduce_mean(axes = var_4840_axes_0, keep_dims = var_4840_keep_dims_0, x = var_4838_cast_fp16)[name = string("op_4840_cast_fp16")];
            fp16 var_4841_to_fp16 = const()[name = string("op_4841_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4842_cast_fp16 = add(x = var_4840_cast_fp16, y = var_4841_to_fp16)[name = string("op_4842_cast_fp16")];
            fp32 var_4843_epsilon_0 = const()[name = string("op_4843_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = rsqrt(epsilon = var_4843_epsilon_0, x = var_4842_cast_fp16)[name = string("op_4843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4844_cast_fp16 = mul(x = x_137_cast_fp16, y = var_4843_cast_fp16)[name = string("op_4844_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487751936)))];
            tensor<fp16, [1, 8, 16, 128]> x_141_cast_fp16 = mul(x = var_4844_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487752256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488800896))))[name = string("model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_4853 = const()[name = string("op_4853"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4854_cast_fp16 = reshape(shape = var_4853, x = linear_79_cast_fp16)[name = string("op_4854_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_45_begin_0 = const()[name = string("first_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_45_end_0 = const()[name = string("first_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_45_end_mask_0 = const()[name = string("first_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_45_cast_fp16 = slice_by_index(begin = first_45_begin_0, end = first_45_end_0, end_mask = first_45_end_mask_0, x = x_139_cast_fp16)[name = string("first_45_cast_fp16")];
            tensor<int32, [4]> second_45_begin_0 = const()[name = string("second_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_45_end_0 = const()[name = string("second_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_45_end_mask_0 = const()[name = string("second_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_45_cast_fp16 = slice_by_index(begin = second_45_begin_0, end = second_45_end_0, end_mask = second_45_end_mask_0, x = x_139_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4868_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4869_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4869_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4871_cast_fp16 = sub(x = var_4868_cast_fp16, y = var_4869_cast_fp16)[name = string("op_4871_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4872_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4872_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4873_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4873_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = add(x = var_4872_cast_fp16, y = var_4873_cast_fp16)[name = string("op_4875_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-1)];
            bool q_23_interleave_0 = const()[name = string("q_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_23_cast_fp16 = concat(axis = var_4877, interleave = q_23_interleave_0, values = (var_4871_cast_fp16, var_4875_cast_fp16))[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> first_47_begin_0 = const()[name = string("first_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_47_end_0 = const()[name = string("first_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_47_end_mask_0 = const()[name = string("first_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_47_cast_fp16 = slice_by_index(begin = first_47_begin_0, end = first_47_end_0, end_mask = first_47_end_mask_0, x = x_141_cast_fp16)[name = string("first_47_cast_fp16")];
            tensor<int32, [4]> second_47_begin_0 = const()[name = string("second_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_47_end_0 = const()[name = string("second_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_47_end_mask_0 = const()[name = string("second_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_47_cast_fp16 = slice_by_index(begin = second_47_begin_0, end = second_47_end_0, end_mask = second_47_end_mask_0, x = x_141_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4889_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4890_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4892_cast_fp16 = sub(x = var_4889_cast_fp16, y = var_4890_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4893_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4894_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = add(x = var_4893_cast_fp16, y = var_4894_cast_fp16)[name = string("op_4896_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(-1)];
            bool k_23_interleave_0 = const()[name = string("k_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_23_cast_fp16 = concat(axis = var_4898, interleave = k_23_interleave_0, values = (var_4892_cast_fp16, var_4896_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_4909_begin_0 = const()[name = string("op_4909_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4909_end_0 = const()[name = string("op_4909_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4909_end_mask_0 = const()[name = string("op_4909_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4909_stride_0 = const()[name = string("op_4909_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4909_cast_fp16 = slice_by_index(begin = var_4909_begin_0, end = var_4909_end_0, end_mask = var_4909_end_mask_0, stride = var_4909_stride_0, x = q_23_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [4]> var_4919_begin_0 = const()[name = string("op_4919_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4919_end_0 = const()[name = string("op_4919_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4919_end_mask_0 = const()[name = string("op_4919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4919_stride_0 = const()[name = string("op_4919_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4919_cast_fp16 = slice_by_index(begin = var_4919_begin_0, end = var_4919_end_0, end_mask = var_4919_end_mask_0, stride = var_4919_stride_0, x = q_23_cast_fp16)[name = string("op_4919_cast_fp16")];
            int32 var_4921 = const()[name = string("op_4921"), val = int32(2)];
            bool stacked_23_interleave_0 = const()[name = string("stacked_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_23_cast_fp16 = concat(axis = var_4921, interleave = stacked_23_interleave_0, values = (var_4909_cast_fp16, var_4919_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = string("op_4927_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = string("op_4927_end_0"), val = tensor<int32, [4]>([12, 8, 4112, 128])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = string("op_4927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4927_cast_fp16 = slice_by_index(begin = var_4927_begin_0, end = var_4927_end_0, end_mask = var_4927_end_mask_0, x = read_state_0)[name = string("op_4927_cast_fp16")];
            tensor<int32, [4]> var_4937_begin_0 = const()[name = string("op_4937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4937_end_0 = const()[name = string("op_4937_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4937_end_mask_0 = const()[name = string("op_4937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4937_cast_fp16 = slice_by_index(begin = var_4937_begin_0, end = var_4937_end_0, end_mask = var_4937_end_mask_0, x = var_4927_cast_fp16)[name = string("op_4937_cast_fp16")];
            bool var_4946_transpose_x_1 = const()[name = string("op_4946_transpose_x_1"), val = bool(false)];
            bool var_4946_transpose_y_1 = const()[name = string("op_4946_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_4946_cast_fp16 = matmul(transpose_x = var_4946_transpose_x_1, transpose_y = var_4946_transpose_y_1, x = var_4937_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4949_perm_0 = const()[name = string("op_4949_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4950_to_fp16 = const()[name = string("op_4950_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_4949_cast_fp16 = transpose(perm = var_4949_perm_0, x = var_4946_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 4096]> var_4951_cast_fp16 = mul(x = var_4949_cast_fp16, y = var_4950_to_fp16)[name = string("op_4951_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_23_cast_fp16 = add(x = var_4951_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_4957_transpose_x_1 = const()[name = string("op_4957_transpose_x_1"), val = bool(false)];
            bool var_4957_transpose_y_1 = const()[name = string("op_4957_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4957_cast_fp16 = matmul(transpose_x = var_4957_transpose_x_1, transpose_y = var_4957_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_4957_cast_fp16")];
            fp16 var_4958_to_fp16 = const()[name = string("op_4958_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4959_cast_fp16 = mul(x = var_4957_cast_fp16, y = var_4958_to_fp16)[name = string("op_4959_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_4959_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_4963 = const()[name = string("op_4963"), val = int32(-1)];
            bool var_4964_interleave_0 = const()[name = string("op_4964_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_4964_cast_fp16 = concat(axis = var_4963, interleave = var_4964_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_4964_cast_fp16")];
            int32 var_4965 = const()[name = string("op_4965"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_23_cast_fp16 = softmax(axis = var_4965, x = var_4964_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_4972_begin_0 = const()[name = string("op_4972_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4972_end_0 = const()[name = string("op_4972_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_4972_end_mask_0 = const()[name = string("op_4972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_4972_cast_fp16 = slice_by_index(begin = var_4972_begin_0, end = var_4972_end_0, end_mask = var_4972_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<int32, [4]> var_4977_begin_0 = const()[name = string("op_4977_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4977_end_0 = const()[name = string("op_4977_end_0"), val = tensor<int32, [4]>([12, 8, 4112, 128])];
            tensor<bool, [4]> var_4977_end_mask_0 = const()[name = string("op_4977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_4977_cast_fp16 = slice_by_index(begin = var_4977_begin_0, end = var_4977_end_0, end_mask = var_4977_end_mask_0, x = read_state_1)[name = string("op_4977_cast_fp16")];
            tensor<int32, [4]> var_4987_begin_0 = const()[name = string("op_4987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4987_end_0 = const()[name = string("op_4987_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_4987_end_mask_0 = const()[name = string("op_4987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_4987_cast_fp16 = slice_by_index(begin = var_4987_begin_0, end = var_4987_end_0, end_mask = var_4987_end_mask_0, x = var_4977_cast_fp16)[name = string("op_4987_cast_fp16")];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = var_4972_cast_fp16, y = var_4987_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<int32, [4]> var_4998_begin_0 = const()[name = string("op_4998_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_4998_end_0 = const()[name = string("op_4998_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_4998_end_mask_0 = const()[name = string("op_4998_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4998_cast_fp16 = slice_by_index(begin = var_4998_begin_0, end = var_4998_end_0, end_mask = var_4998_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4998_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_4854_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4998_cast_fp16, y = v_23_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_4993_cast_fp16, y = var_4999_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5016_begin_0 = const()[name = string("op_5016_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5016_end_0 = const()[name = string("op_5016_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5016_end_mask_0 = const()[name = string("op_5016_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5016_cast_fp16 = slice_by_index(begin = var_5016_begin_0, end = var_5016_end_0, end_mask = var_5016_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5016_cast_fp16")];
            tensor<int32, [4]> var_5031_begin_0 = const()[name = string("op_5031_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5031_end_0 = const()[name = string("op_5031_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5031_end_mask_0 = const()[name = string("op_5031_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5031_cast_fp16 = slice_by_index(begin = var_5031_begin_0, end = var_5031_end_0, end_mask = var_5031_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5031_cast_fp16")];
            int32 var_5033 = const()[name = string("op_5033"), val = int32(1)];
            bool mixed_47_interleave_0 = const()[name = string("mixed_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_47_cast_fp16 = concat(axis = var_5033, interleave = mixed_47_interleave_0, values = (var_5016_cast_fp16, var_5031_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5037_perm_0 = const()[name = string("op_5037_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5041 = const()[name = string("op_5041"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5037_cast_fp16 = transpose(perm = var_5037_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5041, x = var_5037_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488803008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490900224))))[name = string("model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = add(x = x_133_cast_fp16, y = linear_80_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_5052_promoted_to_fp16 = const()[name = string("op_5052_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5054_cast_fp16 = pow(x = x_143_cast_fp16, y = var_5052_promoted_to_fp16)[name = string("op_5054_cast_fp16")];
            tensor<int32, [1]> var_5056_axes_0 = const()[name = string("op_5056_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5056_keep_dims_0 = const()[name = string("op_5056_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5056_cast_fp16 = reduce_mean(axes = var_5056_axes_0, keep_dims = var_5056_keep_dims_0, x = var_5054_cast_fp16)[name = string("op_5056_cast_fp16")];
            fp16 var_5057_to_fp16 = const()[name = string("op_5057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5058_cast_fp16 = add(x = var_5056_cast_fp16, y = var_5057_to_fp16)[name = string("op_5058_cast_fp16")];
            fp32 var_5059_epsilon_0 = const()[name = string("op_5059_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5059_cast_fp16 = rsqrt(epsilon = var_5059_epsilon_0, x = var_5058_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = mul(x = x_143_cast_fp16, y = var_5059_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490902336)))];
            tensor<fp16, [1, 16, 1024]> input_115_cast_fp16 = mul(x = var_5060_cast_fp16, y = model_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490904448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494050240))))[name = string("model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5065_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494056448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497202240))))[name = string("model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_119_cast_fp16 = mul(x = var_5065_cast_fp16, y = linear_82_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497208448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500354240))))[name = string("model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_145_cast_fp16 = add(x = x_143_cast_fp16, y = linear_83_cast_fp16)[name = string("x_145_cast_fp16")];
            fp16 var_5079_promoted_to_fp16 = const()[name = string("op_5079_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5081_cast_fp16 = pow(x = x_145_cast_fp16, y = var_5079_promoted_to_fp16)[name = string("op_5081_cast_fp16")];
            tensor<int32, [1]> var_5083_axes_0 = const()[name = string("op_5083_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5083_keep_dims_0 = const()[name = string("op_5083_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5083_cast_fp16 = reduce_mean(axes = var_5083_axes_0, keep_dims = var_5083_keep_dims_0, x = var_5081_cast_fp16)[name = string("op_5083_cast_fp16")];
            fp16 var_5084_to_fp16 = const()[name = string("op_5084_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5085_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5084_to_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5086_cast_fp16 = rsqrt(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5087_cast_fp16 = mul(x = x_145_cast_fp16, y = var_5086_cast_fp16)[name = string("op_5087_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500356352)))];
            tensor<fp16, [1, 16, 1024]> input_121_cast_fp16 = mul(x = var_5087_cast_fp16, y = model_layers_12_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500358464))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502455680))))[name = string("model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5096 = const()[name = string("op_5096"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5097_cast_fp16 = reshape(shape = var_5096, x = linear_84_cast_fp16)[name = string("op_5097_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5105_promoted_to_fp16 = const()[name = string("op_5105_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_5097_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5107_cast_fp16 = pow(x = x_147_cast_fp16, y = var_5105_promoted_to_fp16)[name = string("op_5107_cast_fp16")];
            tensor<int32, [1]> var_5109_axes_0 = const()[name = string("op_5109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5109_keep_dims_0 = const()[name = string("op_5109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5109_cast_fp16 = reduce_mean(axes = var_5109_axes_0, keep_dims = var_5109_keep_dims_0, x = var_5107_cast_fp16)[name = string("op_5109_cast_fp16")];
            fp16 var_5110_to_fp16 = const()[name = string("op_5110_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5111_cast_fp16 = add(x = var_5109_cast_fp16, y = var_5110_to_fp16)[name = string("op_5111_cast_fp16")];
            fp32 var_5112_epsilon_0 = const()[name = string("op_5112_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5112_cast_fp16 = rsqrt(epsilon = var_5112_epsilon_0, x = var_5111_cast_fp16)[name = string("op_5112_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5113_cast_fp16 = mul(x = x_147_cast_fp16, y = var_5112_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502459840)))];
            tensor<fp16, [1, 16, 16, 128]> x_151_cast_fp16 = mul(x = var_5113_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502460160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503508800))))[name = string("model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5122 = const()[name = string("op_5122"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5123_cast_fp16 = reshape(shape = var_5122, x = linear_85_cast_fp16)[name = string("op_5123_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5131_promoted_to_fp16 = const()[name = string("op_5131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_5123_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5133_cast_fp16 = pow(x = x_149_cast_fp16, y = var_5131_promoted_to_fp16)[name = string("op_5133_cast_fp16")];
            tensor<int32, [1]> var_5135_axes_0 = const()[name = string("op_5135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5135_keep_dims_0 = const()[name = string("op_5135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5135_cast_fp16 = reduce_mean(axes = var_5135_axes_0, keep_dims = var_5135_keep_dims_0, x = var_5133_cast_fp16)[name = string("op_5135_cast_fp16")];
            fp16 var_5136_to_fp16 = const()[name = string("op_5136_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5137_cast_fp16 = add(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5138_cast_fp16 = rsqrt(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5139_cast_fp16 = mul(x = x_149_cast_fp16, y = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503510912)))];
            tensor<fp16, [1, 8, 16, 128]> x_153_cast_fp16 = mul(x = var_5139_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503511232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504559872))))[name = string("model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5148 = const()[name = string("op_5148"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5149_cast_fp16 = reshape(shape = var_5148, x = linear_86_cast_fp16)[name = string("op_5149_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_49_begin_0 = const()[name = string("first_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_49_end_0 = const()[name = string("first_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_49_end_mask_0 = const()[name = string("first_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_49_cast_fp16 = slice_by_index(begin = first_49_begin_0, end = first_49_end_0, end_mask = first_49_end_mask_0, x = x_151_cast_fp16)[name = string("first_49_cast_fp16")];
            tensor<int32, [4]> second_49_begin_0 = const()[name = string("second_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_49_end_0 = const()[name = string("second_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_49_end_mask_0 = const()[name = string("second_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_49_cast_fp16 = slice_by_index(begin = second_49_begin_0, end = second_49_end_0, end_mask = second_49_end_mask_0, x = x_151_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5163_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5164_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5164_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5166_cast_fp16 = sub(x = var_5163_cast_fp16, y = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5167_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5168_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5170_cast_fp16 = add(x = var_5167_cast_fp16, y = var_5168_cast_fp16)[name = string("op_5170_cast_fp16")];
            int32 var_5172 = const()[name = string("op_5172"), val = int32(-1)];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_25_cast_fp16 = concat(axis = var_5172, interleave = q_25_interleave_0, values = (var_5166_cast_fp16, var_5170_cast_fp16))[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> first_51_begin_0 = const()[name = string("first_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_51_end_0 = const()[name = string("first_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_51_end_mask_0 = const()[name = string("first_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_51_cast_fp16 = slice_by_index(begin = first_51_begin_0, end = first_51_end_0, end_mask = first_51_end_mask_0, x = x_153_cast_fp16)[name = string("first_51_cast_fp16")];
            tensor<int32, [4]> second_51_begin_0 = const()[name = string("second_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_51_end_0 = const()[name = string("second_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_51_end_mask_0 = const()[name = string("second_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_51_cast_fp16 = slice_by_index(begin = second_51_begin_0, end = second_51_end_0, end_mask = second_51_end_mask_0, x = x_153_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5184_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5185_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5185_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5187_cast_fp16 = sub(x = var_5184_cast_fp16, y = var_5185_cast_fp16)[name = string("op_5187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5188_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5189_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5191_cast_fp16 = add(x = var_5188_cast_fp16, y = var_5189_cast_fp16)[name = string("op_5191_cast_fp16")];
            int32 var_5193 = const()[name = string("op_5193"), val = int32(-1)];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_25_cast_fp16 = concat(axis = var_5193, interleave = k_25_interleave_0, values = (var_5187_cast_fp16, var_5191_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5204_begin_0 = const()[name = string("op_5204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5204_end_0 = const()[name = string("op_5204_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5204_end_mask_0 = const()[name = string("op_5204_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5204_stride_0 = const()[name = string("op_5204_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5204_cast_fp16 = slice_by_index(begin = var_5204_begin_0, end = var_5204_end_0, end_mask = var_5204_end_mask_0, stride = var_5204_stride_0, x = q_25_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5214_stride_0 = const()[name = string("op_5214_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, stride = var_5214_stride_0, x = q_25_cast_fp16)[name = string("op_5214_cast_fp16")];
            int32 var_5216 = const()[name = string("op_5216"), val = int32(2)];
            bool stacked_25_interleave_0 = const()[name = string("stacked_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_25_cast_fp16 = concat(axis = var_5216, interleave = stacked_25_interleave_0, values = (var_5204_cast_fp16, var_5214_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5222_begin_0 = const()[name = string("op_5222_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5222_end_0 = const()[name = string("op_5222_end_0"), val = tensor<int32, [4]>([13, 8, 4112, 128])];
            tensor<bool, [4]> var_5222_end_mask_0 = const()[name = string("op_5222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, x = read_state_0)[name = string("op_5222_cast_fp16")];
            tensor<int32, [4]> var_5232_begin_0 = const()[name = string("op_5232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5232_end_0 = const()[name = string("op_5232_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5232_end_mask_0 = const()[name = string("op_5232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5232_cast_fp16 = slice_by_index(begin = var_5232_begin_0, end = var_5232_end_0, end_mask = var_5232_end_mask_0, x = var_5222_cast_fp16)[name = string("op_5232_cast_fp16")];
            bool var_5241_transpose_x_1 = const()[name = string("op_5241_transpose_x_1"), val = bool(false)];
            bool var_5241_transpose_y_1 = const()[name = string("op_5241_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5241_cast_fp16 = matmul(transpose_x = var_5241_transpose_x_1, transpose_y = var_5241_transpose_y_1, x = var_5232_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5241_cast_fp16")];
            tensor<int32, [4]> var_5244_perm_0 = const()[name = string("op_5244_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5245_to_fp16 = const()[name = string("op_5245_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5244_cast_fp16 = transpose(perm = var_5244_perm_0, x = var_5241_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 4096]> var_5246_cast_fp16 = mul(x = var_5244_cast_fp16, y = var_5245_to_fp16)[name = string("op_5246_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_25_cast_fp16 = add(x = var_5246_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5252_transpose_x_1 = const()[name = string("op_5252_transpose_x_1"), val = bool(false)];
            bool var_5252_transpose_y_1 = const()[name = string("op_5252_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5252_cast_fp16 = matmul(transpose_x = var_5252_transpose_x_1, transpose_y = var_5252_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5252_cast_fp16")];
            fp16 var_5253_to_fp16 = const()[name = string("op_5253_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5254_cast_fp16 = mul(x = var_5252_cast_fp16, y = var_5253_to_fp16)[name = string("op_5254_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5254_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5258 = const()[name = string("op_5258"), val = int32(-1)];
            bool var_5259_interleave_0 = const()[name = string("op_5259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5259_cast_fp16 = concat(axis = var_5258, interleave = var_5259_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5259_cast_fp16")];
            int32 var_5260 = const()[name = string("op_5260"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_25_cast_fp16 = softmax(axis = var_5260, x = var_5259_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5267_begin_0 = const()[name = string("op_5267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5267_end_0 = const()[name = string("op_5267_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5267_end_mask_0 = const()[name = string("op_5267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5267_cast_fp16 = slice_by_index(begin = var_5267_begin_0, end = var_5267_end_0, end_mask = var_5267_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<int32, [4]> var_5272_begin_0 = const()[name = string("op_5272_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5272_end_0 = const()[name = string("op_5272_end_0"), val = tensor<int32, [4]>([13, 8, 4112, 128])];
            tensor<bool, [4]> var_5272_end_mask_0 = const()[name = string("op_5272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5272_cast_fp16 = slice_by_index(begin = var_5272_begin_0, end = var_5272_end_0, end_mask = var_5272_end_mask_0, x = read_state_1)[name = string("op_5272_cast_fp16")];
            tensor<int32, [4]> var_5282_begin_0 = const()[name = string("op_5282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5282_end_0 = const()[name = string("op_5282_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5282_end_mask_0 = const()[name = string("op_5282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5282_cast_fp16 = slice_by_index(begin = var_5282_begin_0, end = var_5282_end_0, end_mask = var_5282_end_mask_0, x = var_5272_cast_fp16)[name = string("op_5282_cast_fp16")];
            bool var_5288_transpose_x_0 = const()[name = string("op_5288_transpose_x_0"), val = bool(false)];
            bool var_5288_transpose_y_0 = const()[name = string("op_5288_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5288_cast_fp16 = matmul(transpose_x = var_5288_transpose_x_0, transpose_y = var_5288_transpose_y_0, x = var_5267_cast_fp16, y = var_5282_cast_fp16)[name = string("op_5288_cast_fp16")];
            tensor<int32, [4]> var_5293_begin_0 = const()[name = string("op_5293_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5293_end_0 = const()[name = string("op_5293_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5293_end_mask_0 = const()[name = string("op_5293_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5293_cast_fp16 = slice_by_index(begin = var_5293_begin_0, end = var_5293_end_0, end_mask = var_5293_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5293_cast_fp16")];
            bool var_5294_transpose_x_0 = const()[name = string("op_5294_transpose_x_0"), val = bool(false)];
            bool var_5294_transpose_y_0 = const()[name = string("op_5294_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5149_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5294_cast_fp16 = matmul(transpose_x = var_5294_transpose_x_0, transpose_y = var_5294_transpose_y_0, x = var_5293_cast_fp16, y = v_25_cast_fp16)[name = string("op_5294_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5288_cast_fp16, y = var_5294_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5311_begin_0 = const()[name = string("op_5311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5311_end_0 = const()[name = string("op_5311_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5311_end_mask_0 = const()[name = string("op_5311_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5311_cast_fp16 = slice_by_index(begin = var_5311_begin_0, end = var_5311_end_0, end_mask = var_5311_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> var_5326_begin_0 = const()[name = string("op_5326_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5326_end_0 = const()[name = string("op_5326_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5326_end_mask_0 = const()[name = string("op_5326_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5326_cast_fp16 = slice_by_index(begin = var_5326_begin_0, end = var_5326_end_0, end_mask = var_5326_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5326_cast_fp16")];
            int32 var_5328 = const()[name = string("op_5328"), val = int32(1)];
            bool mixed_51_interleave_0 = const()[name = string("mixed_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_51_cast_fp16 = concat(axis = var_5328, interleave = mixed_51_interleave_0, values = (var_5311_cast_fp16, var_5326_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504561984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506659200))))[name = string("model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = linear_87_cast_fp16)[name = string("x_155_cast_fp16")];
            fp16 var_5347_promoted_to_fp16 = const()[name = string("op_5347_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5349_cast_fp16 = pow(x = x_155_cast_fp16, y = var_5347_promoted_to_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5351_axes_0 = const()[name = string("op_5351_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5351_keep_dims_0 = const()[name = string("op_5351_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5351_cast_fp16 = reduce_mean(axes = var_5351_axes_0, keep_dims = var_5351_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5351_cast_fp16")];
            fp16 var_5352_to_fp16 = const()[name = string("op_5352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5353_cast_fp16 = add(x = var_5351_cast_fp16, y = var_5352_to_fp16)[name = string("op_5353_cast_fp16")];
            fp32 var_5354_epsilon_0 = const()[name = string("op_5354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5354_cast_fp16 = rsqrt(epsilon = var_5354_epsilon_0, x = var_5353_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5355_cast_fp16 = mul(x = x_155_cast_fp16, y = var_5354_cast_fp16)[name = string("op_5355_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506661312)))];
            tensor<fp16, [1, 16, 1024]> input_125_cast_fp16 = mul(x = var_5355_cast_fp16, y = model_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506663424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509809216))))[name = string("model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5360_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509815424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512961216))))[name = string("model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_129_cast_fp16 = mul(x = var_5360_cast_fp16, y = linear_89_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512967424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516113216))))[name = string("model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_157_cast_fp16 = add(x = x_155_cast_fp16, y = linear_90_cast_fp16)[name = string("x_157_cast_fp16")];
            fp16 var_5374_promoted_to_fp16 = const()[name = string("op_5374_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5376_cast_fp16 = pow(x = x_157_cast_fp16, y = var_5374_promoted_to_fp16)[name = string("op_5376_cast_fp16")];
            tensor<int32, [1]> var_5378_axes_0 = const()[name = string("op_5378_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5378_keep_dims_0 = const()[name = string("op_5378_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5378_cast_fp16 = reduce_mean(axes = var_5378_axes_0, keep_dims = var_5378_keep_dims_0, x = var_5376_cast_fp16)[name = string("op_5378_cast_fp16")];
            fp16 var_5379_to_fp16 = const()[name = string("op_5379_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5380_cast_fp16 = add(x = var_5378_cast_fp16, y = var_5379_to_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5381_cast_fp16 = rsqrt(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5382_cast_fp16 = mul(x = x_157_cast_fp16, y = var_5381_cast_fp16)[name = string("op_5382_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516115328)))];
            tensor<fp16, [1, 16, 1024]> input_131_cast_fp16 = mul(x = var_5382_cast_fp16, y = model_layers_13_input_layernorm_weight_to_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516117440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518214656))))[name = string("model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5391 = const()[name = string("op_5391"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5392_cast_fp16 = reshape(shape = var_5391, x = linear_91_cast_fp16)[name = string("op_5392_cast_fp16")];
            tensor<int32, [4]> x_159_perm_0 = const()[name = string("x_159_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5400_promoted_to_fp16 = const()[name = string("op_5400_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_159_cast_fp16 = transpose(perm = x_159_perm_0, x = var_5392_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5402_cast_fp16 = pow(x = x_159_cast_fp16, y = var_5400_promoted_to_fp16)[name = string("op_5402_cast_fp16")];
            tensor<int32, [1]> var_5404_axes_0 = const()[name = string("op_5404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5404_keep_dims_0 = const()[name = string("op_5404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5404_cast_fp16 = reduce_mean(axes = var_5404_axes_0, keep_dims = var_5404_keep_dims_0, x = var_5402_cast_fp16)[name = string("op_5404_cast_fp16")];
            fp16 var_5405_to_fp16 = const()[name = string("op_5405_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5406_cast_fp16 = add(x = var_5404_cast_fp16, y = var_5405_to_fp16)[name = string("op_5406_cast_fp16")];
            fp32 var_5407_epsilon_0 = const()[name = string("op_5407_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5407_cast_fp16 = rsqrt(epsilon = var_5407_epsilon_0, x = var_5406_cast_fp16)[name = string("op_5407_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5408_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5407_cast_fp16)[name = string("op_5408_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518218816)))];
            tensor<fp16, [1, 16, 16, 128]> x_163_cast_fp16 = mul(x = var_5408_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518219136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519267776))))[name = string("model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5417 = const()[name = string("op_5417"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5418_cast_fp16 = reshape(shape = var_5417, x = linear_92_cast_fp16)[name = string("op_5418_cast_fp16")];
            tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5426_promoted_to_fp16 = const()[name = string("op_5426_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_5418_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5428_cast_fp16 = pow(x = x_161_cast_fp16, y = var_5426_promoted_to_fp16)[name = string("op_5428_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5428_cast_fp16)[name = string("op_5430_cast_fp16")];
            fp16 var_5431_to_fp16 = const()[name = string("op_5431_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5432_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_to_fp16)[name = string("op_5432_cast_fp16")];
            fp32 var_5433_epsilon_0 = const()[name = string("op_5433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5433_cast_fp16 = rsqrt(epsilon = var_5433_epsilon_0, x = var_5432_cast_fp16)[name = string("op_5433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5434_cast_fp16 = mul(x = x_161_cast_fp16, y = var_5433_cast_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519269888)))];
            tensor<fp16, [1, 8, 16, 128]> x_165_cast_fp16 = mul(x = var_5434_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519270208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520318848))))[name = string("model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5443 = const()[name = string("op_5443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5444_cast_fp16 = reshape(shape = var_5443, x = linear_93_cast_fp16)[name = string("op_5444_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_53_begin_0 = const()[name = string("first_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_53_end_0 = const()[name = string("first_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_53_end_mask_0 = const()[name = string("first_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_53_cast_fp16 = slice_by_index(begin = first_53_begin_0, end = first_53_end_0, end_mask = first_53_end_mask_0, x = x_163_cast_fp16)[name = string("first_53_cast_fp16")];
            tensor<int32, [4]> second_53_begin_0 = const()[name = string("second_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_53_end_0 = const()[name = string("second_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_53_end_mask_0 = const()[name = string("second_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_53_cast_fp16 = slice_by_index(begin = second_53_begin_0, end = second_53_end_0, end_mask = second_53_end_mask_0, x = x_163_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5458_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5459_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5461_cast_fp16 = sub(x = var_5458_cast_fp16, y = var_5459_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5462_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5463_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5465_cast_fp16 = add(x = var_5462_cast_fp16, y = var_5463_cast_fp16)[name = string("op_5465_cast_fp16")];
            int32 var_5467 = const()[name = string("op_5467"), val = int32(-1)];
            bool q_27_interleave_0 = const()[name = string("q_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_27_cast_fp16 = concat(axis = var_5467, interleave = q_27_interleave_0, values = (var_5461_cast_fp16, var_5465_cast_fp16))[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> first_55_begin_0 = const()[name = string("first_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_55_end_0 = const()[name = string("first_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_55_end_mask_0 = const()[name = string("first_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_55_cast_fp16 = slice_by_index(begin = first_55_begin_0, end = first_55_end_0, end_mask = first_55_end_mask_0, x = x_165_cast_fp16)[name = string("first_55_cast_fp16")];
            tensor<int32, [4]> second_55_begin_0 = const()[name = string("second_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_55_end_0 = const()[name = string("second_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_55_end_mask_0 = const()[name = string("second_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_55_cast_fp16 = slice_by_index(begin = second_55_begin_0, end = second_55_end_0, end_mask = second_55_end_mask_0, x = x_165_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5479_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5480_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5482_cast_fp16 = sub(x = var_5479_cast_fp16, y = var_5480_cast_fp16)[name = string("op_5482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5483_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5484_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5486_cast_fp16 = add(x = var_5483_cast_fp16, y = var_5484_cast_fp16)[name = string("op_5486_cast_fp16")];
            int32 var_5488 = const()[name = string("op_5488"), val = int32(-1)];
            bool k_27_interleave_0 = const()[name = string("k_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_27_cast_fp16 = concat(axis = var_5488, interleave = k_27_interleave_0, values = (var_5482_cast_fp16, var_5486_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5499_begin_0 = const()[name = string("op_5499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5499_end_0 = const()[name = string("op_5499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5499_end_mask_0 = const()[name = string("op_5499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5499_stride_0 = const()[name = string("op_5499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5499_cast_fp16 = slice_by_index(begin = var_5499_begin_0, end = var_5499_end_0, end_mask = var_5499_end_mask_0, stride = var_5499_stride_0, x = q_27_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<int32, [4]> var_5509_begin_0 = const()[name = string("op_5509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5509_end_0 = const()[name = string("op_5509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5509_end_mask_0 = const()[name = string("op_5509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5509_stride_0 = const()[name = string("op_5509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5509_cast_fp16 = slice_by_index(begin = var_5509_begin_0, end = var_5509_end_0, end_mask = var_5509_end_mask_0, stride = var_5509_stride_0, x = q_27_cast_fp16)[name = string("op_5509_cast_fp16")];
            int32 var_5511 = const()[name = string("op_5511"), val = int32(2)];
            bool stacked_27_interleave_0 = const()[name = string("stacked_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_27_cast_fp16 = concat(axis = var_5511, interleave = stacked_27_interleave_0, values = (var_5499_cast_fp16, var_5509_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5517_begin_0 = const()[name = string("op_5517_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5517_end_0 = const()[name = string("op_5517_end_0"), val = tensor<int32, [4]>([14, 8, 4112, 128])];
            tensor<bool, [4]> var_5517_end_mask_0 = const()[name = string("op_5517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5517_cast_fp16 = slice_by_index(begin = var_5517_begin_0, end = var_5517_end_0, end_mask = var_5517_end_mask_0, x = read_state_0)[name = string("op_5517_cast_fp16")];
            tensor<int32, [4]> var_5527_begin_0 = const()[name = string("op_5527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5527_end_0 = const()[name = string("op_5527_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5527_end_mask_0 = const()[name = string("op_5527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5527_cast_fp16 = slice_by_index(begin = var_5527_begin_0, end = var_5527_end_0, end_mask = var_5527_end_mask_0, x = var_5517_cast_fp16)[name = string("op_5527_cast_fp16")];
            bool var_5536_transpose_x_1 = const()[name = string("op_5536_transpose_x_1"), val = bool(false)];
            bool var_5536_transpose_y_1 = const()[name = string("op_5536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5536_cast_fp16 = matmul(transpose_x = var_5536_transpose_x_1, transpose_y = var_5536_transpose_y_1, x = var_5527_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5536_cast_fp16")];
            tensor<int32, [4]> var_5539_perm_0 = const()[name = string("op_5539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5540_to_fp16 = const()[name = string("op_5540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5539_cast_fp16 = transpose(perm = var_5539_perm_0, x = var_5536_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 4096]> var_5541_cast_fp16 = mul(x = var_5539_cast_fp16, y = var_5540_to_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_27_cast_fp16 = add(x = var_5541_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5547_transpose_x_1 = const()[name = string("op_5547_transpose_x_1"), val = bool(false)];
            bool var_5547_transpose_y_1 = const()[name = string("op_5547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5547_cast_fp16 = matmul(transpose_x = var_5547_transpose_x_1, transpose_y = var_5547_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5547_cast_fp16")];
            fp16 var_5548_to_fp16 = const()[name = string("op_5548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5549_cast_fp16 = mul(x = var_5547_cast_fp16, y = var_5548_to_fp16)[name = string("op_5549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5553 = const()[name = string("op_5553"), val = int32(-1)];
            bool var_5554_interleave_0 = const()[name = string("op_5554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5554_cast_fp16 = concat(axis = var_5553, interleave = var_5554_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5554_cast_fp16")];
            int32 var_5555 = const()[name = string("op_5555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_27_cast_fp16 = softmax(axis = var_5555, x = var_5554_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5562_begin_0 = const()[name = string("op_5562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5562_end_0 = const()[name = string("op_5562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5562_end_mask_0 = const()[name = string("op_5562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5562_cast_fp16 = slice_by_index(begin = var_5562_begin_0, end = var_5562_end_0, end_mask = var_5562_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5562_cast_fp16")];
            tensor<int32, [4]> var_5567_begin_0 = const()[name = string("op_5567_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5567_end_0 = const()[name = string("op_5567_end_0"), val = tensor<int32, [4]>([14, 8, 4112, 128])];
            tensor<bool, [4]> var_5567_end_mask_0 = const()[name = string("op_5567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5567_cast_fp16 = slice_by_index(begin = var_5567_begin_0, end = var_5567_end_0, end_mask = var_5567_end_mask_0, x = read_state_1)[name = string("op_5567_cast_fp16")];
            tensor<int32, [4]> var_5577_begin_0 = const()[name = string("op_5577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5577_end_0 = const()[name = string("op_5577_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5577_end_mask_0 = const()[name = string("op_5577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5577_cast_fp16 = slice_by_index(begin = var_5577_begin_0, end = var_5577_end_0, end_mask = var_5577_end_mask_0, x = var_5567_cast_fp16)[name = string("op_5577_cast_fp16")];
            bool var_5583_transpose_x_0 = const()[name = string("op_5583_transpose_x_0"), val = bool(false)];
            bool var_5583_transpose_y_0 = const()[name = string("op_5583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5583_cast_fp16 = matmul(transpose_x = var_5583_transpose_x_0, transpose_y = var_5583_transpose_y_0, x = var_5562_cast_fp16, y = var_5577_cast_fp16)[name = string("op_5583_cast_fp16")];
            tensor<int32, [4]> var_5588_begin_0 = const()[name = string("op_5588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5588_end_0 = const()[name = string("op_5588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5588_end_mask_0 = const()[name = string("op_5588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5588_cast_fp16 = slice_by_index(begin = var_5588_begin_0, end = var_5588_end_0, end_mask = var_5588_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5588_cast_fp16")];
            bool var_5589_transpose_x_0 = const()[name = string("op_5589_transpose_x_0"), val = bool(false)];
            bool var_5589_transpose_y_0 = const()[name = string("op_5589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5444_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5589_cast_fp16 = matmul(transpose_x = var_5589_transpose_x_0, transpose_y = var_5589_transpose_y_0, x = var_5588_cast_fp16, y = v_27_cast_fp16)[name = string("op_5589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5583_cast_fp16, y = var_5589_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5606_begin_0 = const()[name = string("op_5606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5606_end_0 = const()[name = string("op_5606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5606_end_mask_0 = const()[name = string("op_5606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5606_cast_fp16 = slice_by_index(begin = var_5606_begin_0, end = var_5606_end_0, end_mask = var_5606_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5621_begin_0 = const()[name = string("op_5621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5621_end_0 = const()[name = string("op_5621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5621_end_mask_0 = const()[name = string("op_5621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5621_cast_fp16 = slice_by_index(begin = var_5621_begin_0, end = var_5621_end_0, end_mask = var_5621_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5621_cast_fp16")];
            int32 var_5623 = const()[name = string("op_5623"), val = int32(1)];
            bool mixed_55_interleave_0 = const()[name = string("mixed_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_55_cast_fp16 = concat(axis = var_5623, interleave = mixed_55_interleave_0, values = (var_5606_cast_fp16, var_5621_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5631 = const()[name = string("op_5631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_5631, x = var_5627_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520320960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522418176))))[name = string("model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_167_cast_fp16 = add(x = x_157_cast_fp16, y = linear_94_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_5642_promoted_to_fp16 = const()[name = string("op_5642_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5644_cast_fp16 = pow(x = x_167_cast_fp16, y = var_5642_promoted_to_fp16)[name = string("op_5644_cast_fp16")];
            tensor<int32, [1]> var_5646_axes_0 = const()[name = string("op_5646_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5646_keep_dims_0 = const()[name = string("op_5646_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5646_cast_fp16 = reduce_mean(axes = var_5646_axes_0, keep_dims = var_5646_keep_dims_0, x = var_5644_cast_fp16)[name = string("op_5646_cast_fp16")];
            fp16 var_5647_to_fp16 = const()[name = string("op_5647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5648_cast_fp16 = add(x = var_5646_cast_fp16, y = var_5647_to_fp16)[name = string("op_5648_cast_fp16")];
            fp32 var_5649_epsilon_0 = const()[name = string("op_5649_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5649_cast_fp16 = rsqrt(epsilon = var_5649_epsilon_0, x = var_5648_cast_fp16)[name = string("op_5649_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5650_cast_fp16 = mul(x = x_167_cast_fp16, y = var_5649_cast_fp16)[name = string("op_5650_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522420288)))];
            tensor<fp16, [1, 16, 1024]> input_135_cast_fp16 = mul(x = var_5650_cast_fp16, y = model_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522422400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525568192))))[name = string("model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5655_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_5655_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525574400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528720192))))[name = string("model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_139_cast_fp16 = mul(x = var_5655_cast_fp16, y = linear_96_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528726400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531872192))))[name = string("model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_169_cast_fp16 = add(x = x_167_cast_fp16, y = linear_97_cast_fp16)[name = string("x_169_cast_fp16")];
            fp16 var_5669_promoted_to_fp16 = const()[name = string("op_5669_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5671_cast_fp16 = pow(x = x_169_cast_fp16, y = var_5669_promoted_to_fp16)[name = string("op_5671_cast_fp16")];
            tensor<int32, [1]> var_5673_axes_0 = const()[name = string("op_5673_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5673_keep_dims_0 = const()[name = string("op_5673_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5673_cast_fp16 = reduce_mean(axes = var_5673_axes_0, keep_dims = var_5673_keep_dims_0, x = var_5671_cast_fp16)[name = string("op_5673_cast_fp16")];
            fp16 var_5674_to_fp16 = const()[name = string("op_5674_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5675_cast_fp16 = add(x = var_5673_cast_fp16, y = var_5674_to_fp16)[name = string("op_5675_cast_fp16")];
            fp32 var_5676_epsilon_0 = const()[name = string("op_5676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5676_cast_fp16 = rsqrt(epsilon = var_5676_epsilon_0, x = var_5675_cast_fp16)[name = string("op_5676_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5677_cast_fp16 = mul(x = x_169_cast_fp16, y = var_5676_cast_fp16)[name = string("op_5677_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531874304)))];
            tensor<fp16, [1, 16, 1024]> input_141_cast_fp16 = mul(x = var_5677_cast_fp16, y = model_layers_14_input_layernorm_weight_to_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531876416))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533973632))))[name = string("model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_5686 = const()[name = string("op_5686"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5687_cast_fp16 = reshape(shape = var_5686, x = linear_98_cast_fp16)[name = string("op_5687_cast_fp16")];
            tensor<int32, [4]> x_171_perm_0 = const()[name = string("x_171_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5695_promoted_to_fp16 = const()[name = string("op_5695_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_171_cast_fp16 = transpose(perm = x_171_perm_0, x = var_5687_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_5697_cast_fp16 = pow(x = x_171_cast_fp16, y = var_5695_promoted_to_fp16)[name = string("op_5697_cast_fp16")];
            tensor<int32, [1]> var_5699_axes_0 = const()[name = string("op_5699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5699_keep_dims_0 = const()[name = string("op_5699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5699_cast_fp16 = reduce_mean(axes = var_5699_axes_0, keep_dims = var_5699_keep_dims_0, x = var_5697_cast_fp16)[name = string("op_5699_cast_fp16")];
            fp16 var_5700_to_fp16 = const()[name = string("op_5700_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5701_cast_fp16 = add(x = var_5699_cast_fp16, y = var_5700_to_fp16)[name = string("op_5701_cast_fp16")];
            fp32 var_5702_epsilon_0 = const()[name = string("op_5702_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5702_cast_fp16 = rsqrt(epsilon = var_5702_epsilon_0, x = var_5701_cast_fp16)[name = string("op_5702_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5703_cast_fp16 = mul(x = x_171_cast_fp16, y = var_5702_cast_fp16)[name = string("op_5703_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533977792)))];
            tensor<fp16, [1, 16, 16, 128]> x_175_cast_fp16 = mul(x = var_5703_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_175_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533978112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535026752))))[name = string("model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_5712 = const()[name = string("op_5712"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5713_cast_fp16 = reshape(shape = var_5712, x = linear_99_cast_fp16)[name = string("op_5713_cast_fp16")];
            tensor<int32, [4]> x_173_perm_0 = const()[name = string("x_173_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5721_promoted_to_fp16 = const()[name = string("op_5721_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_173_cast_fp16 = transpose(perm = x_173_perm_0, x = var_5713_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_5723_cast_fp16 = pow(x = x_173_cast_fp16, y = var_5721_promoted_to_fp16)[name = string("op_5723_cast_fp16")];
            tensor<int32, [1]> var_5725_axes_0 = const()[name = string("op_5725_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5725_keep_dims_0 = const()[name = string("op_5725_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5725_cast_fp16 = reduce_mean(axes = var_5725_axes_0, keep_dims = var_5725_keep_dims_0, x = var_5723_cast_fp16)[name = string("op_5725_cast_fp16")];
            fp16 var_5726_to_fp16 = const()[name = string("op_5726_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5727_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5726_to_fp16)[name = string("op_5727_cast_fp16")];
            fp32 var_5728_epsilon_0 = const()[name = string("op_5728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5728_cast_fp16 = rsqrt(epsilon = var_5728_epsilon_0, x = var_5727_cast_fp16)[name = string("op_5728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5729_cast_fp16 = mul(x = x_173_cast_fp16, y = var_5728_cast_fp16)[name = string("op_5729_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535028864)))];
            tensor<fp16, [1, 8, 16, 128]> x_177_cast_fp16 = mul(x = var_5729_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535029184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536077824))))[name = string("model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5738 = const()[name = string("op_5738"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5739_cast_fp16 = reshape(shape = var_5738, x = linear_100_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_57_begin_0 = const()[name = string("first_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_57_end_0 = const()[name = string("first_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_57_end_mask_0 = const()[name = string("first_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_57_cast_fp16 = slice_by_index(begin = first_57_begin_0, end = first_57_end_0, end_mask = first_57_end_mask_0, x = x_175_cast_fp16)[name = string("first_57_cast_fp16")];
            tensor<int32, [4]> second_57_begin_0 = const()[name = string("second_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_57_end_0 = const()[name = string("second_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_57_end_mask_0 = const()[name = string("second_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_57_cast_fp16 = slice_by_index(begin = second_57_begin_0, end = second_57_end_0, end_mask = second_57_end_mask_0, x = x_175_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5753_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5754_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5756_cast_fp16 = sub(x = var_5753_cast_fp16, y = var_5754_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5757_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5757_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5758_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5758_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5760_cast_fp16 = add(x = var_5757_cast_fp16, y = var_5758_cast_fp16)[name = string("op_5760_cast_fp16")];
            int32 var_5762 = const()[name = string("op_5762"), val = int32(-1)];
            bool q_29_interleave_0 = const()[name = string("q_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_29_cast_fp16 = concat(axis = var_5762, interleave = q_29_interleave_0, values = (var_5756_cast_fp16, var_5760_cast_fp16))[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> first_59_begin_0 = const()[name = string("first_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_59_end_0 = const()[name = string("first_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_59_end_mask_0 = const()[name = string("first_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_59_cast_fp16 = slice_by_index(begin = first_59_begin_0, end = first_59_end_0, end_mask = first_59_end_mask_0, x = x_177_cast_fp16)[name = string("first_59_cast_fp16")];
            tensor<int32, [4]> second_59_begin_0 = const()[name = string("second_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_59_end_0 = const()[name = string("second_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_59_end_mask_0 = const()[name = string("second_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_59_cast_fp16 = slice_by_index(begin = second_59_begin_0, end = second_59_end_0, end_mask = second_59_end_mask_0, x = x_177_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5774_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5775_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5777_cast_fp16 = sub(x = var_5774_cast_fp16, y = var_5775_cast_fp16)[name = string("op_5777_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5778_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5779_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5779_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5781_cast_fp16 = add(x = var_5778_cast_fp16, y = var_5779_cast_fp16)[name = string("op_5781_cast_fp16")];
            int32 var_5783 = const()[name = string("op_5783"), val = int32(-1)];
            bool k_29_interleave_0 = const()[name = string("k_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_29_cast_fp16 = concat(axis = var_5783, interleave = k_29_interleave_0, values = (var_5777_cast_fp16, var_5781_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_5794_begin_0 = const()[name = string("op_5794_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5794_end_0 = const()[name = string("op_5794_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5794_end_mask_0 = const()[name = string("op_5794_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5794_stride_0 = const()[name = string("op_5794_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5794_cast_fp16 = slice_by_index(begin = var_5794_begin_0, end = var_5794_end_0, end_mask = var_5794_end_mask_0, stride = var_5794_stride_0, x = q_29_cast_fp16)[name = string("op_5794_cast_fp16")];
            tensor<int32, [4]> var_5804_begin_0 = const()[name = string("op_5804_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5804_end_0 = const()[name = string("op_5804_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5804_end_mask_0 = const()[name = string("op_5804_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5804_stride_0 = const()[name = string("op_5804_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5804_cast_fp16 = slice_by_index(begin = var_5804_begin_0, end = var_5804_end_0, end_mask = var_5804_end_mask_0, stride = var_5804_stride_0, x = q_29_cast_fp16)[name = string("op_5804_cast_fp16")];
            int32 var_5806 = const()[name = string("op_5806"), val = int32(2)];
            bool stacked_29_interleave_0 = const()[name = string("stacked_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_29_cast_fp16 = concat(axis = var_5806, interleave = stacked_29_interleave_0, values = (var_5794_cast_fp16, var_5804_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_5812_begin_0 = const()[name = string("op_5812_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5812_end_0 = const()[name = string("op_5812_end_0"), val = tensor<int32, [4]>([15, 8, 4112, 128])];
            tensor<bool, [4]> var_5812_end_mask_0 = const()[name = string("op_5812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5812_cast_fp16 = slice_by_index(begin = var_5812_begin_0, end = var_5812_end_0, end_mask = var_5812_end_mask_0, x = read_state_0)[name = string("op_5812_cast_fp16")];
            tensor<int32, [4]> var_5822_begin_0 = const()[name = string("op_5822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5822_end_0 = const()[name = string("op_5822_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5822_end_mask_0 = const()[name = string("op_5822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5822_cast_fp16 = slice_by_index(begin = var_5822_begin_0, end = var_5822_end_0, end_mask = var_5822_end_mask_0, x = var_5812_cast_fp16)[name = string("op_5822_cast_fp16")];
            bool var_5831_transpose_x_1 = const()[name = string("op_5831_transpose_x_1"), val = bool(false)];
            bool var_5831_transpose_y_1 = const()[name = string("op_5831_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_5831_cast_fp16 = matmul(transpose_x = var_5831_transpose_x_1, transpose_y = var_5831_transpose_y_1, x = var_5822_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_5831_cast_fp16")];
            tensor<int32, [4]> var_5834_perm_0 = const()[name = string("op_5834_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5835_to_fp16 = const()[name = string("op_5835_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_5834_cast_fp16 = transpose(perm = var_5834_perm_0, x = var_5831_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 4096]> var_5836_cast_fp16 = mul(x = var_5834_cast_fp16, y = var_5835_to_fp16)[name = string("op_5836_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_29_cast_fp16 = add(x = var_5836_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_5842_transpose_x_1 = const()[name = string("op_5842_transpose_x_1"), val = bool(false)];
            bool var_5842_transpose_y_1 = const()[name = string("op_5842_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5842_cast_fp16 = matmul(transpose_x = var_5842_transpose_x_1, transpose_y = var_5842_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_5842_cast_fp16")];
            fp16 var_5843_to_fp16 = const()[name = string("op_5843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5844_cast_fp16 = mul(x = var_5842_cast_fp16, y = var_5843_to_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_5844_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_5848 = const()[name = string("op_5848"), val = int32(-1)];
            bool var_5849_interleave_0 = const()[name = string("op_5849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_5849_cast_fp16 = concat(axis = var_5848, interleave = var_5849_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_5849_cast_fp16")];
            int32 var_5850 = const()[name = string("op_5850"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_29_cast_fp16 = softmax(axis = var_5850, x = var_5849_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_5857_begin_0 = const()[name = string("op_5857_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5857_end_0 = const()[name = string("op_5857_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_5857_end_mask_0 = const()[name = string("op_5857_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_5857_cast_fp16 = slice_by_index(begin = var_5857_begin_0, end = var_5857_end_0, end_mask = var_5857_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5857_cast_fp16")];
            tensor<int32, [4]> var_5862_begin_0 = const()[name = string("op_5862_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5862_end_0 = const()[name = string("op_5862_end_0"), val = tensor<int32, [4]>([15, 8, 4112, 128])];
            tensor<bool, [4]> var_5862_end_mask_0 = const()[name = string("op_5862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_5862_cast_fp16 = slice_by_index(begin = var_5862_begin_0, end = var_5862_end_0, end_mask = var_5862_end_mask_0, x = read_state_1)[name = string("op_5862_cast_fp16")];
            tensor<int32, [4]> var_5872_begin_0 = const()[name = string("op_5872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5872_end_0 = const()[name = string("op_5872_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_5872_end_mask_0 = const()[name = string("op_5872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_5872_cast_fp16 = slice_by_index(begin = var_5872_begin_0, end = var_5872_end_0, end_mask = var_5872_end_mask_0, x = var_5862_cast_fp16)[name = string("op_5872_cast_fp16")];
            bool var_5878_transpose_x_0 = const()[name = string("op_5878_transpose_x_0"), val = bool(false)];
            bool var_5878_transpose_y_0 = const()[name = string("op_5878_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5878_cast_fp16 = matmul(transpose_x = var_5878_transpose_x_0, transpose_y = var_5878_transpose_y_0, x = var_5857_cast_fp16, y = var_5872_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<int32, [4]> var_5883_begin_0 = const()[name = string("op_5883_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_5883_end_0 = const()[name = string("op_5883_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_5883_end_mask_0 = const()[name = string("op_5883_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5883_cast_fp16 = slice_by_index(begin = var_5883_begin_0, end = var_5883_end_0, end_mask = var_5883_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5883_cast_fp16")];
            bool var_5884_transpose_x_0 = const()[name = string("op_5884_transpose_x_0"), val = bool(false)];
            bool var_5884_transpose_y_0 = const()[name = string("op_5884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_5739_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_5884_cast_fp16 = matmul(transpose_x = var_5884_transpose_x_0, transpose_y = var_5884_transpose_y_0, x = var_5883_cast_fp16, y = v_29_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5884_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_5901_begin_0 = const()[name = string("op_5901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5901_end_0 = const()[name = string("op_5901_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5901_end_mask_0 = const()[name = string("op_5901_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5901_cast_fp16 = slice_by_index(begin = var_5901_begin_0, end = var_5901_end_0, end_mask = var_5901_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5901_cast_fp16")];
            tensor<int32, [4]> var_5916_begin_0 = const()[name = string("op_5916_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5916_end_0 = const()[name = string("op_5916_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5916_end_mask_0 = const()[name = string("op_5916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5916_cast_fp16 = slice_by_index(begin = var_5916_begin_0, end = var_5916_end_0, end_mask = var_5916_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5916_cast_fp16")];
            int32 var_5918 = const()[name = string("op_5918"), val = int32(1)];
            bool mixed_59_interleave_0 = const()[name = string("mixed_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_59_cast_fp16 = concat(axis = var_5918, interleave = mixed_59_interleave_0, values = (var_5901_cast_fp16, var_5916_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_5922_perm_0 = const()[name = string("op_5922_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5926 = const()[name = string("op_5926"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5922_cast_fp16 = transpose(perm = var_5922_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_5926, x = var_5922_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536079936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538177152))))[name = string("model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = add(x = x_169_cast_fp16, y = linear_101_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_5937_promoted_to_fp16 = const()[name = string("op_5937_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5939_cast_fp16 = pow(x = x_179_cast_fp16, y = var_5937_promoted_to_fp16)[name = string("op_5939_cast_fp16")];
            tensor<int32, [1]> var_5941_axes_0 = const()[name = string("op_5941_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5941_keep_dims_0 = const()[name = string("op_5941_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5941_cast_fp16 = reduce_mean(axes = var_5941_axes_0, keep_dims = var_5941_keep_dims_0, x = var_5939_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 var_5942_to_fp16 = const()[name = string("op_5942_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5943_cast_fp16 = add(x = var_5941_cast_fp16, y = var_5942_to_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5944_cast_fp16 = rsqrt(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5945_cast_fp16 = mul(x = x_179_cast_fp16, y = var_5944_cast_fp16)[name = string("op_5945_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538179264)))];
            tensor<fp16, [1, 16, 1024]> input_145_cast_fp16 = mul(x = var_5945_cast_fp16, y = model_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538181376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541327168))))[name = string("model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5950_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_5950_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541333376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544479168))))[name = string("model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_149_cast_fp16 = mul(x = var_5950_cast_fp16, y = linear_103_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544485376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547631168))))[name = string("model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_104_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_181_cast_fp16 = add(x = x_179_cast_fp16, y = linear_104_cast_fp16)[name = string("x_181_cast_fp16")];
            fp16 var_5964_promoted_to_fp16 = const()[name = string("op_5964_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5966_cast_fp16 = pow(x = x_181_cast_fp16, y = var_5964_promoted_to_fp16)[name = string("op_5966_cast_fp16")];
            tensor<int32, [1]> var_5968_axes_0 = const()[name = string("op_5968_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5968_keep_dims_0 = const()[name = string("op_5968_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5968_cast_fp16 = reduce_mean(axes = var_5968_axes_0, keep_dims = var_5968_keep_dims_0, x = var_5966_cast_fp16)[name = string("op_5968_cast_fp16")];
            fp16 var_5969_to_fp16 = const()[name = string("op_5969_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5970_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5969_to_fp16)[name = string("op_5970_cast_fp16")];
            fp32 var_5971_epsilon_0 = const()[name = string("op_5971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5971_cast_fp16 = rsqrt(epsilon = var_5971_epsilon_0, x = var_5970_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5972_cast_fp16 = mul(x = x_181_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5972_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547633280)))];
            tensor<fp16, [1, 16, 1024]> input_151_cast_fp16 = mul(x = var_5972_cast_fp16, y = model_layers_15_input_layernorm_weight_to_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547635392))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549732608))))[name = string("model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_5981 = const()[name = string("op_5981"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5982_cast_fp16 = reshape(shape = var_5981, x = linear_105_cast_fp16)[name = string("op_5982_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5990_promoted_to_fp16 = const()[name = string("op_5990_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5982_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_5992_cast_fp16 = pow(x = x_183_cast_fp16, y = var_5990_promoted_to_fp16)[name = string("op_5992_cast_fp16")];
            tensor<int32, [1]> var_5994_axes_0 = const()[name = string("op_5994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5994_keep_dims_0 = const()[name = string("op_5994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5994_cast_fp16 = reduce_mean(axes = var_5994_axes_0, keep_dims = var_5994_keep_dims_0, x = var_5992_cast_fp16)[name = string("op_5994_cast_fp16")];
            fp16 var_5995_to_fp16 = const()[name = string("op_5995_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5996_cast_fp16 = add(x = var_5994_cast_fp16, y = var_5995_to_fp16)[name = string("op_5996_cast_fp16")];
            fp32 var_5997_epsilon_0 = const()[name = string("op_5997_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5997_cast_fp16 = rsqrt(epsilon = var_5997_epsilon_0, x = var_5996_cast_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5998_cast_fp16 = mul(x = x_183_cast_fp16, y = var_5997_cast_fp16)[name = string("op_5998_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549736768)))];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = mul(x = var_5998_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549737088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550785728))))[name = string("model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6007 = const()[name = string("op_6007"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6008_cast_fp16 = reshape(shape = var_6007, x = linear_106_cast_fp16)[name = string("op_6008_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6016_promoted_to_fp16 = const()[name = string("op_6016_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_6008_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6018_cast_fp16 = pow(x = x_185_cast_fp16, y = var_6016_promoted_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<int32, [1]> var_6020_axes_0 = const()[name = string("op_6020_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6020_keep_dims_0 = const()[name = string("op_6020_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6020_cast_fp16 = reduce_mean(axes = var_6020_axes_0, keep_dims = var_6020_keep_dims_0, x = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp16 var_6021_to_fp16 = const()[name = string("op_6021_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6022_cast_fp16 = add(x = var_6020_cast_fp16, y = var_6021_to_fp16)[name = string("op_6022_cast_fp16")];
            fp32 var_6023_epsilon_0 = const()[name = string("op_6023_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6023_cast_fp16 = rsqrt(epsilon = var_6023_epsilon_0, x = var_6022_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6024_cast_fp16 = mul(x = x_185_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6024_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550787840)))];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = mul(x = var_6024_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550788160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551836800))))[name = string("model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6033 = const()[name = string("op_6033"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6034_cast_fp16 = reshape(shape = var_6033, x = linear_107_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_61_begin_0 = const()[name = string("first_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_61_end_0 = const()[name = string("first_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_61_end_mask_0 = const()[name = string("first_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_61_cast_fp16 = slice_by_index(begin = first_61_begin_0, end = first_61_end_0, end_mask = first_61_end_mask_0, x = x_187_cast_fp16)[name = string("first_61_cast_fp16")];
            tensor<int32, [4]> second_61_begin_0 = const()[name = string("second_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_61_end_0 = const()[name = string("second_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_61_end_mask_0 = const()[name = string("second_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_61_cast_fp16 = slice_by_index(begin = second_61_begin_0, end = second_61_end_0, end_mask = second_61_end_mask_0, x = x_187_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6048_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6049_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6049_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6051_cast_fp16 = sub(x = var_6048_cast_fp16, y = var_6049_cast_fp16)[name = string("op_6051_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6052_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6052_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6053_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6055_cast_fp16 = add(x = var_6052_cast_fp16, y = var_6053_cast_fp16)[name = string("op_6055_cast_fp16")];
            int32 var_6057 = const()[name = string("op_6057"), val = int32(-1)];
            bool q_31_interleave_0 = const()[name = string("q_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_31_cast_fp16 = concat(axis = var_6057, interleave = q_31_interleave_0, values = (var_6051_cast_fp16, var_6055_cast_fp16))[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> first_63_begin_0 = const()[name = string("first_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_63_end_0 = const()[name = string("first_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_63_end_mask_0 = const()[name = string("first_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_63_cast_fp16 = slice_by_index(begin = first_63_begin_0, end = first_63_end_0, end_mask = first_63_end_mask_0, x = x_189_cast_fp16)[name = string("first_63_cast_fp16")];
            tensor<int32, [4]> second_63_begin_0 = const()[name = string("second_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_63_end_0 = const()[name = string("second_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_63_end_mask_0 = const()[name = string("second_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_63_cast_fp16 = slice_by_index(begin = second_63_begin_0, end = second_63_end_0, end_mask = second_63_end_mask_0, x = x_189_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6069_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6070_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6070_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6072_cast_fp16 = sub(x = var_6069_cast_fp16, y = var_6070_cast_fp16)[name = string("op_6072_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6073_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6074_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6076_cast_fp16 = add(x = var_6073_cast_fp16, y = var_6074_cast_fp16)[name = string("op_6076_cast_fp16")];
            int32 var_6078 = const()[name = string("op_6078"), val = int32(-1)];
            bool k_31_interleave_0 = const()[name = string("k_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_31_cast_fp16 = concat(axis = var_6078, interleave = k_31_interleave_0, values = (var_6072_cast_fp16, var_6076_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6089_begin_0 = const()[name = string("op_6089_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6089_end_0 = const()[name = string("op_6089_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6089_end_mask_0 = const()[name = string("op_6089_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6089_stride_0 = const()[name = string("op_6089_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6089_cast_fp16 = slice_by_index(begin = var_6089_begin_0, end = var_6089_end_0, end_mask = var_6089_end_mask_0, stride = var_6089_stride_0, x = q_31_cast_fp16)[name = string("op_6089_cast_fp16")];
            tensor<int32, [4]> var_6099_begin_0 = const()[name = string("op_6099_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6099_end_0 = const()[name = string("op_6099_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6099_end_mask_0 = const()[name = string("op_6099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6099_stride_0 = const()[name = string("op_6099_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6099_cast_fp16 = slice_by_index(begin = var_6099_begin_0, end = var_6099_end_0, end_mask = var_6099_end_mask_0, stride = var_6099_stride_0, x = q_31_cast_fp16)[name = string("op_6099_cast_fp16")];
            int32 var_6101 = const()[name = string("op_6101"), val = int32(2)];
            bool stacked_31_interleave_0 = const()[name = string("stacked_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_31_cast_fp16 = concat(axis = var_6101, interleave = stacked_31_interleave_0, values = (var_6089_cast_fp16, var_6099_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6107_begin_0 = const()[name = string("op_6107_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6107_end_0 = const()[name = string("op_6107_end_0"), val = tensor<int32, [4]>([16, 8, 4112, 128])];
            tensor<bool, [4]> var_6107_end_mask_0 = const()[name = string("op_6107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6107_cast_fp16 = slice_by_index(begin = var_6107_begin_0, end = var_6107_end_0, end_mask = var_6107_end_mask_0, x = read_state_0)[name = string("op_6107_cast_fp16")];
            tensor<int32, [4]> var_6117_begin_0 = const()[name = string("op_6117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6117_end_0 = const()[name = string("op_6117_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6117_end_mask_0 = const()[name = string("op_6117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6117_cast_fp16 = slice_by_index(begin = var_6117_begin_0, end = var_6117_end_0, end_mask = var_6117_end_mask_0, x = var_6107_cast_fp16)[name = string("op_6117_cast_fp16")];
            bool var_6126_transpose_x_1 = const()[name = string("op_6126_transpose_x_1"), val = bool(false)];
            bool var_6126_transpose_y_1 = const()[name = string("op_6126_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6126_cast_fp16 = matmul(transpose_x = var_6126_transpose_x_1, transpose_y = var_6126_transpose_y_1, x = var_6117_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<int32, [4]> var_6129_perm_0 = const()[name = string("op_6129_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6130_to_fp16 = const()[name = string("op_6130_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6129_cast_fp16 = transpose(perm = var_6129_perm_0, x = var_6126_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 4096]> var_6131_cast_fp16 = mul(x = var_6129_cast_fp16, y = var_6130_to_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_31_cast_fp16 = add(x = var_6131_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6137_transpose_x_1 = const()[name = string("op_6137_transpose_x_1"), val = bool(false)];
            bool var_6137_transpose_y_1 = const()[name = string("op_6137_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6137_cast_fp16 = matmul(transpose_x = var_6137_transpose_x_1, transpose_y = var_6137_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6139_cast_fp16 = mul(x = var_6137_cast_fp16, y = var_6138_to_fp16)[name = string("op_6139_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6139_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6143 = const()[name = string("op_6143"), val = int32(-1)];
            bool var_6144_interleave_0 = const()[name = string("op_6144_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6144_cast_fp16 = concat(axis = var_6143, interleave = var_6144_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6144_cast_fp16")];
            int32 var_6145 = const()[name = string("op_6145"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_31_cast_fp16 = softmax(axis = var_6145, x = var_6144_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6152_begin_0 = const()[name = string("op_6152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6152_end_0 = const()[name = string("op_6152_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6152_end_mask_0 = const()[name = string("op_6152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6152_cast_fp16 = slice_by_index(begin = var_6152_begin_0, end = var_6152_end_0, end_mask = var_6152_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<int32, [4]> var_6157_begin_0 = const()[name = string("op_6157_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6157_end_0 = const()[name = string("op_6157_end_0"), val = tensor<int32, [4]>([16, 8, 4112, 128])];
            tensor<bool, [4]> var_6157_end_mask_0 = const()[name = string("op_6157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6157_cast_fp16 = slice_by_index(begin = var_6157_begin_0, end = var_6157_end_0, end_mask = var_6157_end_mask_0, x = read_state_1)[name = string("op_6157_cast_fp16")];
            tensor<int32, [4]> var_6167_begin_0 = const()[name = string("op_6167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6167_end_0 = const()[name = string("op_6167_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6167_end_mask_0 = const()[name = string("op_6167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6167_cast_fp16 = slice_by_index(begin = var_6167_begin_0, end = var_6167_end_0, end_mask = var_6167_end_mask_0, x = var_6157_cast_fp16)[name = string("op_6167_cast_fp16")];
            bool var_6173_transpose_x_0 = const()[name = string("op_6173_transpose_x_0"), val = bool(false)];
            bool var_6173_transpose_y_0 = const()[name = string("op_6173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6173_cast_fp16 = matmul(transpose_x = var_6173_transpose_x_0, transpose_y = var_6173_transpose_y_0, x = var_6152_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6173_cast_fp16")];
            tensor<int32, [4]> var_6178_begin_0 = const()[name = string("op_6178_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6178_end_0 = const()[name = string("op_6178_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6178_end_mask_0 = const()[name = string("op_6178_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6178_cast_fp16 = slice_by_index(begin = var_6178_begin_0, end = var_6178_end_0, end_mask = var_6178_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6178_cast_fp16")];
            bool var_6179_transpose_x_0 = const()[name = string("op_6179_transpose_x_0"), val = bool(false)];
            bool var_6179_transpose_y_0 = const()[name = string("op_6179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6034_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6179_cast_fp16 = matmul(transpose_x = var_6179_transpose_x_0, transpose_y = var_6179_transpose_y_0, x = var_6178_cast_fp16, y = v_31_cast_fp16)[name = string("op_6179_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6173_cast_fp16, y = var_6179_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6196_begin_0 = const()[name = string("op_6196_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6196_end_0 = const()[name = string("op_6196_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6196_end_mask_0 = const()[name = string("op_6196_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6196_cast_fp16 = slice_by_index(begin = var_6196_begin_0, end = var_6196_end_0, end_mask = var_6196_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6196_cast_fp16")];
            tensor<int32, [4]> var_6211_begin_0 = const()[name = string("op_6211_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6211_end_0 = const()[name = string("op_6211_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6211_end_mask_0 = const()[name = string("op_6211_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6211_cast_fp16 = slice_by_index(begin = var_6211_begin_0, end = var_6211_end_0, end_mask = var_6211_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6211_cast_fp16")];
            int32 var_6213 = const()[name = string("op_6213"), val = int32(1)];
            bool mixed_63_interleave_0 = const()[name = string("mixed_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_63_cast_fp16 = concat(axis = var_6213, interleave = mixed_63_interleave_0, values = (var_6196_cast_fp16, var_6211_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6217_perm_0 = const()[name = string("op_6217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6221 = const()[name = string("op_6221"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6217_cast_fp16 = transpose(perm = var_6217_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6221, x = var_6217_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551838912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553936128))))[name = string("model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_108_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = linear_108_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_6232_promoted_to_fp16 = const()[name = string("op_6232_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6234_cast_fp16 = pow(x = x_191_cast_fp16, y = var_6232_promoted_to_fp16)[name = string("op_6234_cast_fp16")];
            tensor<int32, [1]> var_6236_axes_0 = const()[name = string("op_6236_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6236_keep_dims_0 = const()[name = string("op_6236_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6236_cast_fp16 = reduce_mean(axes = var_6236_axes_0, keep_dims = var_6236_keep_dims_0, x = var_6234_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6238_cast_fp16 = add(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            fp32 var_6239_epsilon_0 = const()[name = string("op_6239_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6239_cast_fp16 = rsqrt(epsilon = var_6239_epsilon_0, x = var_6238_cast_fp16)[name = string("op_6239_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6240_cast_fp16 = mul(x = x_191_cast_fp16, y = var_6239_cast_fp16)[name = string("op_6240_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553938240)))];
            tensor<fp16, [1, 16, 1024]> input_155_cast_fp16 = mul(x = var_6240_cast_fp16, y = model_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553940352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557086144))))[name = string("model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6245_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557092352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560238144))))[name = string("model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_159_cast_fp16 = mul(x = var_6245_cast_fp16, y = linear_110_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560244352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563390144))))[name = string("model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_193_cast_fp16 = add(x = x_191_cast_fp16, y = linear_111_cast_fp16)[name = string("x_193_cast_fp16")];
            fp16 var_6259_promoted_to_fp16 = const()[name = string("op_6259_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6261_cast_fp16 = pow(x = x_193_cast_fp16, y = var_6259_promoted_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<int32, [1]> var_6263_axes_0 = const()[name = string("op_6263_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6263_keep_dims_0 = const()[name = string("op_6263_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6263_cast_fp16 = reduce_mean(axes = var_6263_axes_0, keep_dims = var_6263_keep_dims_0, x = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp16 var_6264_to_fp16 = const()[name = string("op_6264_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6265_cast_fp16 = add(x = var_6263_cast_fp16, y = var_6264_to_fp16)[name = string("op_6265_cast_fp16")];
            fp32 var_6266_epsilon_0 = const()[name = string("op_6266_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6266_cast_fp16 = rsqrt(epsilon = var_6266_epsilon_0, x = var_6265_cast_fp16)[name = string("op_6266_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6267_cast_fp16 = mul(x = x_193_cast_fp16, y = var_6266_cast_fp16)[name = string("op_6267_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563392256)))];
            tensor<fp16, [1, 16, 1024]> input_161_cast_fp16 = mul(x = var_6267_cast_fp16, y = model_layers_16_input_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563394368))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565491584))))[name = string("model_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_6276 = const()[name = string("op_6276"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6277_cast_fp16 = reshape(shape = var_6276, x = linear_112_cast_fp16)[name = string("op_6277_cast_fp16")];
            tensor<int32, [4]> x_195_perm_0 = const()[name = string("x_195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6285_promoted_to_fp16 = const()[name = string("op_6285_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_195_cast_fp16 = transpose(perm = x_195_perm_0, x = var_6277_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6287_cast_fp16 = pow(x = x_195_cast_fp16, y = var_6285_promoted_to_fp16)[name = string("op_6287_cast_fp16")];
            tensor<int32, [1]> var_6289_axes_0 = const()[name = string("op_6289_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6289_keep_dims_0 = const()[name = string("op_6289_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6289_cast_fp16 = reduce_mean(axes = var_6289_axes_0, keep_dims = var_6289_keep_dims_0, x = var_6287_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 var_6290_to_fp16 = const()[name = string("op_6290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6291_cast_fp16 = add(x = var_6289_cast_fp16, y = var_6290_to_fp16)[name = string("op_6291_cast_fp16")];
            fp32 var_6292_epsilon_0 = const()[name = string("op_6292_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6292_cast_fp16 = rsqrt(epsilon = var_6292_epsilon_0, x = var_6291_cast_fp16)[name = string("op_6292_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6293_cast_fp16 = mul(x = x_195_cast_fp16, y = var_6292_cast_fp16)[name = string("op_6293_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565495744)))];
            tensor<fp16, [1, 16, 16, 128]> x_199_cast_fp16 = mul(x = var_6293_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565496064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566544704))))[name = string("model_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_6302 = const()[name = string("op_6302"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6303_cast_fp16 = reshape(shape = var_6302, x = linear_113_cast_fp16)[name = string("op_6303_cast_fp16")];
            tensor<int32, [4]> x_197_perm_0 = const()[name = string("x_197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6311_promoted_to_fp16 = const()[name = string("op_6311_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_197_cast_fp16 = transpose(perm = x_197_perm_0, x = var_6303_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6313_cast_fp16 = pow(x = x_197_cast_fp16, y = var_6311_promoted_to_fp16)[name = string("op_6313_cast_fp16")];
            tensor<int32, [1]> var_6315_axes_0 = const()[name = string("op_6315_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6315_keep_dims_0 = const()[name = string("op_6315_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6315_cast_fp16 = reduce_mean(axes = var_6315_axes_0, keep_dims = var_6315_keep_dims_0, x = var_6313_cast_fp16)[name = string("op_6315_cast_fp16")];
            fp16 var_6316_to_fp16 = const()[name = string("op_6316_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6317_cast_fp16 = add(x = var_6315_cast_fp16, y = var_6316_to_fp16)[name = string("op_6317_cast_fp16")];
            fp32 var_6318_epsilon_0 = const()[name = string("op_6318_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6318_cast_fp16 = rsqrt(epsilon = var_6318_epsilon_0, x = var_6317_cast_fp16)[name = string("op_6318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6319_cast_fp16 = mul(x = x_197_cast_fp16, y = var_6318_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566546816)))];
            tensor<fp16, [1, 8, 16, 128]> x_201_cast_fp16 = mul(x = var_6319_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566547136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567595776))))[name = string("model_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_6328 = const()[name = string("op_6328"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6329_cast_fp16 = reshape(shape = var_6328, x = linear_114_cast_fp16)[name = string("op_6329_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_65_begin_0 = const()[name = string("first_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_65_end_0 = const()[name = string("first_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_65_end_mask_0 = const()[name = string("first_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_65_cast_fp16 = slice_by_index(begin = first_65_begin_0, end = first_65_end_0, end_mask = first_65_end_mask_0, x = x_199_cast_fp16)[name = string("first_65_cast_fp16")];
            tensor<int32, [4]> second_65_begin_0 = const()[name = string("second_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_65_end_0 = const()[name = string("second_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_65_end_mask_0 = const()[name = string("second_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_65_cast_fp16 = slice_by_index(begin = second_65_begin_0, end = second_65_end_0, end_mask = second_65_end_mask_0, x = x_199_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6343_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6343_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6344_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6346_cast_fp16 = sub(x = var_6343_cast_fp16, y = var_6344_cast_fp16)[name = string("op_6346_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6347_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6347_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6348_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6350_cast_fp16 = add(x = var_6347_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            int32 var_6352 = const()[name = string("op_6352"), val = int32(-1)];
            bool q_33_interleave_0 = const()[name = string("q_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_33_cast_fp16 = concat(axis = var_6352, interleave = q_33_interleave_0, values = (var_6346_cast_fp16, var_6350_cast_fp16))[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> first_67_begin_0 = const()[name = string("first_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_67_end_0 = const()[name = string("first_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_67_end_mask_0 = const()[name = string("first_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_67_cast_fp16 = slice_by_index(begin = first_67_begin_0, end = first_67_end_0, end_mask = first_67_end_mask_0, x = x_201_cast_fp16)[name = string("first_67_cast_fp16")];
            tensor<int32, [4]> second_67_begin_0 = const()[name = string("second_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_67_end_0 = const()[name = string("second_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_67_end_mask_0 = const()[name = string("second_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_67_cast_fp16 = slice_by_index(begin = second_67_begin_0, end = second_67_end_0, end_mask = second_67_end_mask_0, x = x_201_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6364_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6365_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6365_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6367_cast_fp16 = sub(x = var_6364_cast_fp16, y = var_6365_cast_fp16)[name = string("op_6367_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6368_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6369_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6371_cast_fp16 = add(x = var_6368_cast_fp16, y = var_6369_cast_fp16)[name = string("op_6371_cast_fp16")];
            int32 var_6373 = const()[name = string("op_6373"), val = int32(-1)];
            bool k_33_interleave_0 = const()[name = string("k_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_33_cast_fp16 = concat(axis = var_6373, interleave = k_33_interleave_0, values = (var_6367_cast_fp16, var_6371_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6384_begin_0 = const()[name = string("op_6384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6384_end_0 = const()[name = string("op_6384_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6384_end_mask_0 = const()[name = string("op_6384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6384_stride_0 = const()[name = string("op_6384_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6384_cast_fp16 = slice_by_index(begin = var_6384_begin_0, end = var_6384_end_0, end_mask = var_6384_end_mask_0, stride = var_6384_stride_0, x = q_33_cast_fp16)[name = string("op_6384_cast_fp16")];
            tensor<int32, [4]> var_6394_begin_0 = const()[name = string("op_6394_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6394_end_0 = const()[name = string("op_6394_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6394_end_mask_0 = const()[name = string("op_6394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6394_stride_0 = const()[name = string("op_6394_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6394_cast_fp16 = slice_by_index(begin = var_6394_begin_0, end = var_6394_end_0, end_mask = var_6394_end_mask_0, stride = var_6394_stride_0, x = q_33_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6396 = const()[name = string("op_6396"), val = int32(2)];
            bool stacked_33_interleave_0 = const()[name = string("stacked_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_33_cast_fp16 = concat(axis = var_6396, interleave = stacked_33_interleave_0, values = (var_6384_cast_fp16, var_6394_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6402_begin_0 = const()[name = string("op_6402_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6402_end_0 = const()[name = string("op_6402_end_0"), val = tensor<int32, [4]>([17, 8, 4112, 128])];
            tensor<bool, [4]> var_6402_end_mask_0 = const()[name = string("op_6402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6402_cast_fp16 = slice_by_index(begin = var_6402_begin_0, end = var_6402_end_0, end_mask = var_6402_end_mask_0, x = read_state_0)[name = string("op_6402_cast_fp16")];
            tensor<int32, [4]> var_6412_begin_0 = const()[name = string("op_6412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6412_end_0 = const()[name = string("op_6412_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6412_end_mask_0 = const()[name = string("op_6412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6412_cast_fp16 = slice_by_index(begin = var_6412_begin_0, end = var_6412_end_0, end_mask = var_6412_end_mask_0, x = var_6402_cast_fp16)[name = string("op_6412_cast_fp16")];
            bool var_6421_transpose_x_1 = const()[name = string("op_6421_transpose_x_1"), val = bool(false)];
            bool var_6421_transpose_y_1 = const()[name = string("op_6421_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6421_cast_fp16 = matmul(transpose_x = var_6421_transpose_x_1, transpose_y = var_6421_transpose_y_1, x = var_6412_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<int32, [4]> var_6424_perm_0 = const()[name = string("op_6424_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6425_to_fp16 = const()[name = string("op_6425_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6424_cast_fp16 = transpose(perm = var_6424_perm_0, x = var_6421_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 4096]> var_6426_cast_fp16 = mul(x = var_6424_cast_fp16, y = var_6425_to_fp16)[name = string("op_6426_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_33_cast_fp16 = add(x = var_6426_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6432_transpose_x_1 = const()[name = string("op_6432_transpose_x_1"), val = bool(false)];
            bool var_6432_transpose_y_1 = const()[name = string("op_6432_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6432_cast_fp16 = matmul(transpose_x = var_6432_transpose_x_1, transpose_y = var_6432_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6432_cast_fp16")];
            fp16 var_6433_to_fp16 = const()[name = string("op_6433_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6434_cast_fp16 = mul(x = var_6432_cast_fp16, y = var_6433_to_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6434_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6438 = const()[name = string("op_6438"), val = int32(-1)];
            bool var_6439_interleave_0 = const()[name = string("op_6439_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6439_cast_fp16 = concat(axis = var_6438, interleave = var_6439_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6439_cast_fp16")];
            int32 var_6440 = const()[name = string("op_6440"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_33_cast_fp16 = softmax(axis = var_6440, x = var_6439_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6447_begin_0 = const()[name = string("op_6447_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6447_end_0 = const()[name = string("op_6447_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6447_end_mask_0 = const()[name = string("op_6447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6447_cast_fp16 = slice_by_index(begin = var_6447_begin_0, end = var_6447_end_0, end_mask = var_6447_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6447_cast_fp16")];
            tensor<int32, [4]> var_6452_begin_0 = const()[name = string("op_6452_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6452_end_0 = const()[name = string("op_6452_end_0"), val = tensor<int32, [4]>([17, 8, 4112, 128])];
            tensor<bool, [4]> var_6452_end_mask_0 = const()[name = string("op_6452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6452_cast_fp16 = slice_by_index(begin = var_6452_begin_0, end = var_6452_end_0, end_mask = var_6452_end_mask_0, x = read_state_1)[name = string("op_6452_cast_fp16")];
            tensor<int32, [4]> var_6462_begin_0 = const()[name = string("op_6462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6462_end_0 = const()[name = string("op_6462_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6462_end_mask_0 = const()[name = string("op_6462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6462_cast_fp16 = slice_by_index(begin = var_6462_begin_0, end = var_6462_end_0, end_mask = var_6462_end_mask_0, x = var_6452_cast_fp16)[name = string("op_6462_cast_fp16")];
            bool var_6468_transpose_x_0 = const()[name = string("op_6468_transpose_x_0"), val = bool(false)];
            bool var_6468_transpose_y_0 = const()[name = string("op_6468_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6468_cast_fp16 = matmul(transpose_x = var_6468_transpose_x_0, transpose_y = var_6468_transpose_y_0, x = var_6447_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6468_cast_fp16")];
            tensor<int32, [4]> var_6473_begin_0 = const()[name = string("op_6473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6473_end_0 = const()[name = string("op_6473_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6473_end_mask_0 = const()[name = string("op_6473_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6473_cast_fp16 = slice_by_index(begin = var_6473_begin_0, end = var_6473_end_0, end_mask = var_6473_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6473_cast_fp16")];
            bool var_6474_transpose_x_0 = const()[name = string("op_6474_transpose_x_0"), val = bool(false)];
            bool var_6474_transpose_y_0 = const()[name = string("op_6474_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6329_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6474_cast_fp16 = matmul(transpose_x = var_6474_transpose_x_0, transpose_y = var_6474_transpose_y_0, x = var_6473_cast_fp16, y = v_33_cast_fp16)[name = string("op_6474_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6468_cast_fp16, y = var_6474_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6491_begin_0 = const()[name = string("op_6491_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6491_end_0 = const()[name = string("op_6491_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6491_end_mask_0 = const()[name = string("op_6491_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6491_cast_fp16 = slice_by_index(begin = var_6491_begin_0, end = var_6491_end_0, end_mask = var_6491_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6491_cast_fp16")];
            tensor<int32, [4]> var_6506_begin_0 = const()[name = string("op_6506_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6506_end_0 = const()[name = string("op_6506_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6506_end_mask_0 = const()[name = string("op_6506_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6506_cast_fp16 = slice_by_index(begin = var_6506_begin_0, end = var_6506_end_0, end_mask = var_6506_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6506_cast_fp16")];
            int32 var_6508 = const()[name = string("op_6508"), val = int32(1)];
            bool mixed_67_interleave_0 = const()[name = string("mixed_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_67_cast_fp16 = concat(axis = var_6508, interleave = mixed_67_interleave_0, values = (var_6491_cast_fp16, var_6506_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6512_perm_0 = const()[name = string("op_6512_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6516 = const()[name = string("op_6516"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6512_cast_fp16 = transpose(perm = var_6512_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6516, x = var_6512_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567597888))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569695104))))[name = string("model_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = add(x = x_193_cast_fp16, y = linear_115_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_6527_promoted_to_fp16 = const()[name = string("op_6527_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6529_cast_fp16 = pow(x = x_203_cast_fp16, y = var_6527_promoted_to_fp16)[name = string("op_6529_cast_fp16")];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6531_keep_dims_0 = const()[name = string("op_6531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6531_cast_fp16 = reduce_mean(axes = var_6531_axes_0, keep_dims = var_6531_keep_dims_0, x = var_6529_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6533_cast_fp16 = add(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("op_6533_cast_fp16")];
            fp32 var_6534_epsilon_0 = const()[name = string("op_6534_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6534_cast_fp16 = rsqrt(epsilon = var_6534_epsilon_0, x = var_6533_cast_fp16)[name = string("op_6534_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6535_cast_fp16 = mul(x = x_203_cast_fp16, y = var_6534_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569697216)))];
            tensor<fp16, [1, 16, 1024]> input_165_cast_fp16 = mul(x = var_6535_cast_fp16, y = model_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569699328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572845120))))[name = string("model_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6540_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572851328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575997120))))[name = string("model_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_169_cast_fp16 = mul(x = var_6540_cast_fp16, y = linear_117_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576003328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579149120))))[name = string("model_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_205_cast_fp16 = add(x = x_203_cast_fp16, y = linear_118_cast_fp16)[name = string("x_205_cast_fp16")];
            fp16 var_6554_promoted_to_fp16 = const()[name = string("op_6554_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6556_cast_fp16 = pow(x = x_205_cast_fp16, y = var_6554_promoted_to_fp16)[name = string("op_6556_cast_fp16")];
            tensor<int32, [1]> var_6558_axes_0 = const()[name = string("op_6558_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6558_keep_dims_0 = const()[name = string("op_6558_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6558_cast_fp16 = reduce_mean(axes = var_6558_axes_0, keep_dims = var_6558_keep_dims_0, x = var_6556_cast_fp16)[name = string("op_6558_cast_fp16")];
            fp16 var_6559_to_fp16 = const()[name = string("op_6559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6560_cast_fp16 = add(x = var_6558_cast_fp16, y = var_6559_to_fp16)[name = string("op_6560_cast_fp16")];
            fp32 var_6561_epsilon_0 = const()[name = string("op_6561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6561_cast_fp16 = rsqrt(epsilon = var_6561_epsilon_0, x = var_6560_cast_fp16)[name = string("op_6561_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6562_cast_fp16 = mul(x = x_205_cast_fp16, y = var_6561_cast_fp16)[name = string("op_6562_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579151232)))];
            tensor<fp16, [1, 16, 1024]> input_171_cast_fp16 = mul(x = var_6562_cast_fp16, y = model_layers_17_input_layernorm_weight_to_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579153344))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581250560))))[name = string("model_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6571 = const()[name = string("op_6571"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6572_cast_fp16 = reshape(shape = var_6571, x = linear_119_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<int32, [4]> x_207_perm_0 = const()[name = string("x_207_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6580_promoted_to_fp16 = const()[name = string("op_6580_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = transpose(perm = x_207_perm_0, x = var_6572_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_6582_cast_fp16 = pow(x = x_207_cast_fp16, y = var_6580_promoted_to_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6584_axes_0 = const()[name = string("op_6584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6584_keep_dims_0 = const()[name = string("op_6584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6584_cast_fp16 = reduce_mean(axes = var_6584_axes_0, keep_dims = var_6584_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6584_cast_fp16")];
            fp16 var_6585_to_fp16 = const()[name = string("op_6585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6586_cast_fp16 = add(x = var_6584_cast_fp16, y = var_6585_to_fp16)[name = string("op_6586_cast_fp16")];
            fp32 var_6587_epsilon_0 = const()[name = string("op_6587_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6587_cast_fp16 = rsqrt(epsilon = var_6587_epsilon_0, x = var_6586_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6588_cast_fp16 = mul(x = x_207_cast_fp16, y = var_6587_cast_fp16)[name = string("op_6588_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581254720)))];
            tensor<fp16, [1, 16, 16, 128]> x_211_cast_fp16 = mul(x = var_6588_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_211_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581255040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582303680))))[name = string("model_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6597 = const()[name = string("op_6597"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6598_cast_fp16 = reshape(shape = var_6597, x = linear_120_cast_fp16)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6606_promoted_to_fp16 = const()[name = string("op_6606_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_6598_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_6608_cast_fp16 = pow(x = x_209_cast_fp16, y = var_6606_promoted_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<int32, [1]> var_6610_axes_0 = const()[name = string("op_6610_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6610_keep_dims_0 = const()[name = string("op_6610_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6610_cast_fp16 = reduce_mean(axes = var_6610_axes_0, keep_dims = var_6610_keep_dims_0, x = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp16 var_6611_to_fp16 = const()[name = string("op_6611_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6612_cast_fp16 = add(x = var_6610_cast_fp16, y = var_6611_to_fp16)[name = string("op_6612_cast_fp16")];
            fp32 var_6613_epsilon_0 = const()[name = string("op_6613_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6613_cast_fp16 = rsqrt(epsilon = var_6613_epsilon_0, x = var_6612_cast_fp16)[name = string("op_6613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6614_cast_fp16 = mul(x = x_209_cast_fp16, y = var_6613_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582305792)))];
            tensor<fp16, [1, 8, 16, 128]> x_213_cast_fp16 = mul(x = var_6614_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582306112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583354752))))[name = string("model_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_6623 = const()[name = string("op_6623"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6624_cast_fp16 = reshape(shape = var_6623, x = linear_121_cast_fp16)[name = string("op_6624_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_69_begin_0 = const()[name = string("first_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_69_end_0 = const()[name = string("first_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_69_end_mask_0 = const()[name = string("first_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_69_cast_fp16 = slice_by_index(begin = first_69_begin_0, end = first_69_end_0, end_mask = first_69_end_mask_0, x = x_211_cast_fp16)[name = string("first_69_cast_fp16")];
            tensor<int32, [4]> second_69_begin_0 = const()[name = string("second_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_69_end_0 = const()[name = string("second_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_69_end_mask_0 = const()[name = string("second_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_69_cast_fp16 = slice_by_index(begin = second_69_begin_0, end = second_69_end_0, end_mask = second_69_end_mask_0, x = x_211_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6638_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6638_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6639_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6639_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6641_cast_fp16 = sub(x = var_6638_cast_fp16, y = var_6639_cast_fp16)[name = string("op_6641_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6642_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6642_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6643_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6645_cast_fp16 = add(x = var_6642_cast_fp16, y = var_6643_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6647 = const()[name = string("op_6647"), val = int32(-1)];
            bool q_35_interleave_0 = const()[name = string("q_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_35_cast_fp16 = concat(axis = var_6647, interleave = q_35_interleave_0, values = (var_6641_cast_fp16, var_6645_cast_fp16))[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> first_71_begin_0 = const()[name = string("first_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_71_end_0 = const()[name = string("first_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_71_end_mask_0 = const()[name = string("first_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_71_cast_fp16 = slice_by_index(begin = first_71_begin_0, end = first_71_end_0, end_mask = first_71_end_mask_0, x = x_213_cast_fp16)[name = string("first_71_cast_fp16")];
            tensor<int32, [4]> second_71_begin_0 = const()[name = string("second_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_71_end_0 = const()[name = string("second_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_71_end_mask_0 = const()[name = string("second_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_71_cast_fp16 = slice_by_index(begin = second_71_begin_0, end = second_71_end_0, end_mask = second_71_end_mask_0, x = x_213_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6659_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6659_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6660_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6660_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6662_cast_fp16 = sub(x = var_6659_cast_fp16, y = var_6660_cast_fp16)[name = string("op_6662_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6663_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6664_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6664_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6666_cast_fp16 = add(x = var_6663_cast_fp16, y = var_6664_cast_fp16)[name = string("op_6666_cast_fp16")];
            int32 var_6668 = const()[name = string("op_6668"), val = int32(-1)];
            bool k_35_interleave_0 = const()[name = string("k_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_35_cast_fp16 = concat(axis = var_6668, interleave = k_35_interleave_0, values = (var_6662_cast_fp16, var_6666_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_6679_begin_0 = const()[name = string("op_6679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6679_end_0 = const()[name = string("op_6679_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6679_end_mask_0 = const()[name = string("op_6679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6679_stride_0 = const()[name = string("op_6679_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6679_cast_fp16 = slice_by_index(begin = var_6679_begin_0, end = var_6679_end_0, end_mask = var_6679_end_mask_0, stride = var_6679_stride_0, x = q_35_cast_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [4]> var_6689_begin_0 = const()[name = string("op_6689_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6689_end_0 = const()[name = string("op_6689_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6689_end_mask_0 = const()[name = string("op_6689_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6689_stride_0 = const()[name = string("op_6689_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6689_cast_fp16 = slice_by_index(begin = var_6689_begin_0, end = var_6689_end_0, end_mask = var_6689_end_mask_0, stride = var_6689_stride_0, x = q_35_cast_fp16)[name = string("op_6689_cast_fp16")];
            int32 var_6691 = const()[name = string("op_6691"), val = int32(2)];
            bool stacked_35_interleave_0 = const()[name = string("stacked_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_35_cast_fp16 = concat(axis = var_6691, interleave = stacked_35_interleave_0, values = (var_6679_cast_fp16, var_6689_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([18, 8, 4112, 128])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = read_state_0)[name = string("op_6697_cast_fp16")];
            tensor<int32, [4]> var_6707_begin_0 = const()[name = string("op_6707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6707_end_0 = const()[name = string("op_6707_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6707_end_mask_0 = const()[name = string("op_6707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6707_cast_fp16 = slice_by_index(begin = var_6707_begin_0, end = var_6707_end_0, end_mask = var_6707_end_mask_0, x = var_6697_cast_fp16)[name = string("op_6707_cast_fp16")];
            bool var_6716_transpose_x_1 = const()[name = string("op_6716_transpose_x_1"), val = bool(false)];
            bool var_6716_transpose_y_1 = const()[name = string("op_6716_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_6716_cast_fp16 = matmul(transpose_x = var_6716_transpose_x_1, transpose_y = var_6716_transpose_y_1, x = var_6707_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_6716_cast_fp16")];
            tensor<int32, [4]> var_6719_perm_0 = const()[name = string("op_6719_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6720_to_fp16 = const()[name = string("op_6720_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_6719_cast_fp16 = transpose(perm = var_6719_perm_0, x = var_6716_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 4096]> var_6721_cast_fp16 = mul(x = var_6719_cast_fp16, y = var_6720_to_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_35_cast_fp16 = add(x = var_6721_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_6727_transpose_x_1 = const()[name = string("op_6727_transpose_x_1"), val = bool(false)];
            bool var_6727_transpose_y_1 = const()[name = string("op_6727_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6727_cast_fp16 = matmul(transpose_x = var_6727_transpose_x_1, transpose_y = var_6727_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_6727_cast_fp16")];
            fp16 var_6728_to_fp16 = const()[name = string("op_6728_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6729_cast_fp16 = mul(x = var_6727_cast_fp16, y = var_6728_to_fp16)[name = string("op_6729_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_6729_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_6733 = const()[name = string("op_6733"), val = int32(-1)];
            bool var_6734_interleave_0 = const()[name = string("op_6734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_6734_cast_fp16 = concat(axis = var_6733, interleave = var_6734_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_6734_cast_fp16")];
            int32 var_6735 = const()[name = string("op_6735"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_35_cast_fp16 = softmax(axis = var_6735, x = var_6734_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_6742_begin_0 = const()[name = string("op_6742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6742_end_0 = const()[name = string("op_6742_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_6742_end_mask_0 = const()[name = string("op_6742_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_6742_cast_fp16 = slice_by_index(begin = var_6742_begin_0, end = var_6742_end_0, end_mask = var_6742_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [4]> var_6747_begin_0 = const()[name = string("op_6747_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6747_end_0 = const()[name = string("op_6747_end_0"), val = tensor<int32, [4]>([18, 8, 4112, 128])];
            tensor<bool, [4]> var_6747_end_mask_0 = const()[name = string("op_6747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6747_cast_fp16 = slice_by_index(begin = var_6747_begin_0, end = var_6747_end_0, end_mask = var_6747_end_mask_0, x = read_state_1)[name = string("op_6747_cast_fp16")];
            tensor<int32, [4]> var_6757_begin_0 = const()[name = string("op_6757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6757_end_0 = const()[name = string("op_6757_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_6757_end_mask_0 = const()[name = string("op_6757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_6757_cast_fp16 = slice_by_index(begin = var_6757_begin_0, end = var_6757_end_0, end_mask = var_6757_end_mask_0, x = var_6747_cast_fp16)[name = string("op_6757_cast_fp16")];
            bool var_6763_transpose_x_0 = const()[name = string("op_6763_transpose_x_0"), val = bool(false)];
            bool var_6763_transpose_y_0 = const()[name = string("op_6763_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6763_cast_fp16 = matmul(transpose_x = var_6763_transpose_x_0, transpose_y = var_6763_transpose_y_0, x = var_6742_cast_fp16, y = var_6757_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<int32, [4]> var_6768_begin_0 = const()[name = string("op_6768_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_6768_end_0 = const()[name = string("op_6768_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_6768_end_mask_0 = const()[name = string("op_6768_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6768_cast_fp16 = slice_by_index(begin = var_6768_begin_0, end = var_6768_end_0, end_mask = var_6768_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6768_cast_fp16")];
            bool var_6769_transpose_x_0 = const()[name = string("op_6769_transpose_x_0"), val = bool(false)];
            bool var_6769_transpose_y_0 = const()[name = string("op_6769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_6624_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_6769_cast_fp16 = matmul(transpose_x = var_6769_transpose_x_0, transpose_y = var_6769_transpose_y_0, x = var_6768_cast_fp16, y = v_35_cast_fp16)[name = string("op_6769_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_6763_cast_fp16, y = var_6769_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_6786_begin_0 = const()[name = string("op_6786_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6786_end_0 = const()[name = string("op_6786_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6786_end_mask_0 = const()[name = string("op_6786_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6786_cast_fp16 = slice_by_index(begin = var_6786_begin_0, end = var_6786_end_0, end_mask = var_6786_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<int32, [4]> var_6801_begin_0 = const()[name = string("op_6801_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6801_end_0 = const()[name = string("op_6801_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6801_end_mask_0 = const()[name = string("op_6801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6801_cast_fp16 = slice_by_index(begin = var_6801_begin_0, end = var_6801_end_0, end_mask = var_6801_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6801_cast_fp16")];
            int32 var_6803 = const()[name = string("op_6803"), val = int32(1)];
            bool mixed_71_interleave_0 = const()[name = string("mixed_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_71_cast_fp16 = concat(axis = var_6803, interleave = mixed_71_interleave_0, values = (var_6786_cast_fp16, var_6801_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_6807_perm_0 = const()[name = string("op_6807_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6811 = const()[name = string("op_6811"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6807_cast_fp16 = transpose(perm = var_6807_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_6811, x = var_6807_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583356864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585454080))))[name = string("model_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_122_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_215_cast_fp16 = add(x = x_205_cast_fp16, y = linear_122_cast_fp16)[name = string("x_215_cast_fp16")];
            fp16 var_6822_promoted_to_fp16 = const()[name = string("op_6822_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6824_cast_fp16 = pow(x = x_215_cast_fp16, y = var_6822_promoted_to_fp16)[name = string("op_6824_cast_fp16")];
            tensor<int32, [1]> var_6826_axes_0 = const()[name = string("op_6826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6826_keep_dims_0 = const()[name = string("op_6826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6826_cast_fp16 = reduce_mean(axes = var_6826_axes_0, keep_dims = var_6826_keep_dims_0, x = var_6824_cast_fp16)[name = string("op_6826_cast_fp16")];
            fp16 var_6827_to_fp16 = const()[name = string("op_6827_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6828_cast_fp16 = add(x = var_6826_cast_fp16, y = var_6827_to_fp16)[name = string("op_6828_cast_fp16")];
            fp32 var_6829_epsilon_0 = const()[name = string("op_6829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6829_cast_fp16 = rsqrt(epsilon = var_6829_epsilon_0, x = var_6828_cast_fp16)[name = string("op_6829_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6830_cast_fp16 = mul(x = x_215_cast_fp16, y = var_6829_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585456192)))];
            tensor<fp16, [1, 16, 1024]> input_175_cast_fp16 = mul(x = var_6830_cast_fp16, y = model_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585458304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588604096))))[name = string("model_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6835_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_6835_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588610304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591756096))))[name = string("model_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_179_cast_fp16 = mul(x = var_6835_cast_fp16, y = linear_124_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591762304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594908096))))[name = string("model_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_217_cast_fp16 = add(x = x_215_cast_fp16, y = linear_125_cast_fp16)[name = string("x_217_cast_fp16")];
            fp16 var_6849_promoted_to_fp16 = const()[name = string("op_6849_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6851_cast_fp16 = pow(x = x_217_cast_fp16, y = var_6849_promoted_to_fp16)[name = string("op_6851_cast_fp16")];
            tensor<int32, [1]> var_6853_axes_0 = const()[name = string("op_6853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6853_keep_dims_0 = const()[name = string("op_6853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6853_cast_fp16 = reduce_mean(axes = var_6853_axes_0, keep_dims = var_6853_keep_dims_0, x = var_6851_cast_fp16)[name = string("op_6853_cast_fp16")];
            fp16 var_6854_to_fp16 = const()[name = string("op_6854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6855_cast_fp16 = add(x = var_6853_cast_fp16, y = var_6854_to_fp16)[name = string("op_6855_cast_fp16")];
            fp32 var_6856_epsilon_0 = const()[name = string("op_6856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6856_cast_fp16 = rsqrt(epsilon = var_6856_epsilon_0, x = var_6855_cast_fp16)[name = string("op_6856_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6857_cast_fp16 = mul(x = x_217_cast_fp16, y = var_6856_cast_fp16)[name = string("op_6857_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594910208)))];
            tensor<fp16, [1, 16, 1024]> input_181_cast_fp16 = mul(x = var_6857_cast_fp16, y = model_layers_18_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594912320))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597009536))))[name = string("model_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_6866 = const()[name = string("op_6866"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6867_cast_fp16 = reshape(shape = var_6866, x = linear_126_cast_fp16)[name = string("op_6867_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6875_promoted_to_fp16 = const()[name = string("op_6875_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6867_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_6877_cast_fp16 = pow(x = x_219_cast_fp16, y = var_6875_promoted_to_fp16)[name = string("op_6877_cast_fp16")];
            tensor<int32, [1]> var_6879_axes_0 = const()[name = string("op_6879_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6879_keep_dims_0 = const()[name = string("op_6879_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6879_cast_fp16 = reduce_mean(axes = var_6879_axes_0, keep_dims = var_6879_keep_dims_0, x = var_6877_cast_fp16)[name = string("op_6879_cast_fp16")];
            fp16 var_6880_to_fp16 = const()[name = string("op_6880_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6881_cast_fp16 = add(x = var_6879_cast_fp16, y = var_6880_to_fp16)[name = string("op_6881_cast_fp16")];
            fp32 var_6882_epsilon_0 = const()[name = string("op_6882_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6882_cast_fp16 = rsqrt(epsilon = var_6882_epsilon_0, x = var_6881_cast_fp16)[name = string("op_6882_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6883_cast_fp16 = mul(x = x_219_cast_fp16, y = var_6882_cast_fp16)[name = string("op_6883_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597013696)))];
            tensor<fp16, [1, 16, 16, 128]> x_223_cast_fp16 = mul(x = var_6883_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597014016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598062656))))[name = string("model_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6892 = const()[name = string("op_6892"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6893_cast_fp16 = reshape(shape = var_6892, x = linear_127_cast_fp16)[name = string("op_6893_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6901_promoted_to_fp16 = const()[name = string("op_6901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6893_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_6903_cast_fp16 = pow(x = x_221_cast_fp16, y = var_6901_promoted_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<int32, [1]> var_6905_axes_0 = const()[name = string("op_6905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6905_keep_dims_0 = const()[name = string("op_6905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6905_cast_fp16 = reduce_mean(axes = var_6905_axes_0, keep_dims = var_6905_keep_dims_0, x = var_6903_cast_fp16)[name = string("op_6905_cast_fp16")];
            fp16 var_6906_to_fp16 = const()[name = string("op_6906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6907_cast_fp16 = add(x = var_6905_cast_fp16, y = var_6906_to_fp16)[name = string("op_6907_cast_fp16")];
            fp32 var_6908_epsilon_0 = const()[name = string("op_6908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6908_cast_fp16 = rsqrt(epsilon = var_6908_epsilon_0, x = var_6907_cast_fp16)[name = string("op_6908_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6909_cast_fp16 = mul(x = x_221_cast_fp16, y = var_6908_cast_fp16)[name = string("op_6909_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598064768)))];
            tensor<fp16, [1, 8, 16, 128]> x_225_cast_fp16 = mul(x = var_6909_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598065088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599113728))))[name = string("model_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6918 = const()[name = string("op_6918"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6919_cast_fp16 = reshape(shape = var_6918, x = linear_128_cast_fp16)[name = string("op_6919_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_73_begin_0 = const()[name = string("first_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_73_end_0 = const()[name = string("first_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_73_end_mask_0 = const()[name = string("first_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_73_cast_fp16 = slice_by_index(begin = first_73_begin_0, end = first_73_end_0, end_mask = first_73_end_mask_0, x = x_223_cast_fp16)[name = string("first_73_cast_fp16")];
            tensor<int32, [4]> second_73_begin_0 = const()[name = string("second_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_73_end_0 = const()[name = string("second_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_73_end_mask_0 = const()[name = string("second_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_73_cast_fp16 = slice_by_index(begin = second_73_begin_0, end = second_73_end_0, end_mask = second_73_end_mask_0, x = x_223_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6933_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6933_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6934_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6934_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6936_cast_fp16 = sub(x = var_6933_cast_fp16, y = var_6934_cast_fp16)[name = string("op_6936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6937_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6938_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6938_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6940_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6938_cast_fp16)[name = string("op_6940_cast_fp16")];
            int32 var_6942 = const()[name = string("op_6942"), val = int32(-1)];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_37_cast_fp16 = concat(axis = var_6942, interleave = q_37_interleave_0, values = (var_6936_cast_fp16, var_6940_cast_fp16))[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> first_75_begin_0 = const()[name = string("first_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_75_end_0 = const()[name = string("first_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_75_end_mask_0 = const()[name = string("first_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_75_cast_fp16 = slice_by_index(begin = first_75_begin_0, end = first_75_end_0, end_mask = first_75_end_mask_0, x = x_225_cast_fp16)[name = string("first_75_cast_fp16")];
            tensor<int32, [4]> second_75_begin_0 = const()[name = string("second_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_75_end_0 = const()[name = string("second_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_75_end_mask_0 = const()[name = string("second_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_75_cast_fp16 = slice_by_index(begin = second_75_begin_0, end = second_75_end_0, end_mask = second_75_end_mask_0, x = x_225_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6954_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6954_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6955_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6955_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6957_cast_fp16 = sub(x = var_6954_cast_fp16, y = var_6955_cast_fp16)[name = string("op_6957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6958_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6959_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6961_cast_fp16 = add(x = var_6958_cast_fp16, y = var_6959_cast_fp16)[name = string("op_6961_cast_fp16")];
            int32 var_6963 = const()[name = string("op_6963"), val = int32(-1)];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_37_cast_fp16 = concat(axis = var_6963, interleave = k_37_interleave_0, values = (var_6957_cast_fp16, var_6961_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_6974_begin_0 = const()[name = string("op_6974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6974_end_0 = const()[name = string("op_6974_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6974_end_mask_0 = const()[name = string("op_6974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6974_stride_0 = const()[name = string("op_6974_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6974_cast_fp16 = slice_by_index(begin = var_6974_begin_0, end = var_6974_end_0, end_mask = var_6974_end_mask_0, stride = var_6974_stride_0, x = q_37_cast_fp16)[name = string("op_6974_cast_fp16")];
            tensor<int32, [4]> var_6984_begin_0 = const()[name = string("op_6984_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6984_end_0 = const()[name = string("op_6984_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6984_end_mask_0 = const()[name = string("op_6984_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6984_stride_0 = const()[name = string("op_6984_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6984_cast_fp16 = slice_by_index(begin = var_6984_begin_0, end = var_6984_end_0, end_mask = var_6984_end_mask_0, stride = var_6984_stride_0, x = q_37_cast_fp16)[name = string("op_6984_cast_fp16")];
            int32 var_6986 = const()[name = string("op_6986"), val = int32(2)];
            bool stacked_37_interleave_0 = const()[name = string("stacked_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_37_cast_fp16 = concat(axis = var_6986, interleave = stacked_37_interleave_0, values = (var_6974_cast_fp16, var_6984_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_6992_begin_0 = const()[name = string("op_6992_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_6992_end_0 = const()[name = string("op_6992_end_0"), val = tensor<int32, [4]>([19, 8, 4112, 128])];
            tensor<bool, [4]> var_6992_end_mask_0 = const()[name = string("op_6992_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_6992_cast_fp16 = slice_by_index(begin = var_6992_begin_0, end = var_6992_end_0, end_mask = var_6992_end_mask_0, x = read_state_0)[name = string("op_6992_cast_fp16")];
            tensor<int32, [4]> var_7002_begin_0 = const()[name = string("op_7002_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7002_end_0 = const()[name = string("op_7002_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7002_end_mask_0 = const()[name = string("op_7002_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7002_cast_fp16 = slice_by_index(begin = var_7002_begin_0, end = var_7002_end_0, end_mask = var_7002_end_mask_0, x = var_6992_cast_fp16)[name = string("op_7002_cast_fp16")];
            bool var_7011_transpose_x_1 = const()[name = string("op_7011_transpose_x_1"), val = bool(false)];
            bool var_7011_transpose_y_1 = const()[name = string("op_7011_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7011_cast_fp16 = matmul(transpose_x = var_7011_transpose_x_1, transpose_y = var_7011_transpose_y_1, x = var_7002_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<int32, [4]> var_7014_perm_0 = const()[name = string("op_7014_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7015_to_fp16 = const()[name = string("op_7015_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7014_cast_fp16 = transpose(perm = var_7014_perm_0, x = var_7011_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 4096]> var_7016_cast_fp16 = mul(x = var_7014_cast_fp16, y = var_7015_to_fp16)[name = string("op_7016_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_37_cast_fp16 = add(x = var_7016_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7022_transpose_x_1 = const()[name = string("op_7022_transpose_x_1"), val = bool(false)];
            bool var_7022_transpose_y_1 = const()[name = string("op_7022_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7022_cast_fp16 = matmul(transpose_x = var_7022_transpose_x_1, transpose_y = var_7022_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7022_cast_fp16")];
            fp16 var_7023_to_fp16 = const()[name = string("op_7023_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7024_cast_fp16 = mul(x = var_7022_cast_fp16, y = var_7023_to_fp16)[name = string("op_7024_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7024_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7028 = const()[name = string("op_7028"), val = int32(-1)];
            bool var_7029_interleave_0 = const()[name = string("op_7029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7029_cast_fp16 = concat(axis = var_7028, interleave = var_7029_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7029_cast_fp16")];
            int32 var_7030 = const()[name = string("op_7030"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_37_cast_fp16 = softmax(axis = var_7030, x = var_7029_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7037_begin_0 = const()[name = string("op_7037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7037_end_0 = const()[name = string("op_7037_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7037_end_mask_0 = const()[name = string("op_7037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7037_cast_fp16 = slice_by_index(begin = var_7037_begin_0, end = var_7037_end_0, end_mask = var_7037_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7037_cast_fp16")];
            tensor<int32, [4]> var_7042_begin_0 = const()[name = string("op_7042_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7042_end_0 = const()[name = string("op_7042_end_0"), val = tensor<int32, [4]>([19, 8, 4112, 128])];
            tensor<bool, [4]> var_7042_end_mask_0 = const()[name = string("op_7042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7042_cast_fp16 = slice_by_index(begin = var_7042_begin_0, end = var_7042_end_0, end_mask = var_7042_end_mask_0, x = read_state_1)[name = string("op_7042_cast_fp16")];
            tensor<int32, [4]> var_7052_begin_0 = const()[name = string("op_7052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7052_end_0 = const()[name = string("op_7052_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7052_end_mask_0 = const()[name = string("op_7052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7052_cast_fp16 = slice_by_index(begin = var_7052_begin_0, end = var_7052_end_0, end_mask = var_7052_end_mask_0, x = var_7042_cast_fp16)[name = string("op_7052_cast_fp16")];
            bool var_7058_transpose_x_0 = const()[name = string("op_7058_transpose_x_0"), val = bool(false)];
            bool var_7058_transpose_y_0 = const()[name = string("op_7058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7058_cast_fp16 = matmul(transpose_x = var_7058_transpose_x_0, transpose_y = var_7058_transpose_y_0, x = var_7037_cast_fp16, y = var_7052_cast_fp16)[name = string("op_7058_cast_fp16")];
            tensor<int32, [4]> var_7063_begin_0 = const()[name = string("op_7063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7063_end_0 = const()[name = string("op_7063_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7063_end_mask_0 = const()[name = string("op_7063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7063_cast_fp16 = slice_by_index(begin = var_7063_begin_0, end = var_7063_end_0, end_mask = var_7063_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7063_cast_fp16")];
            bool var_7064_transpose_x_0 = const()[name = string("op_7064_transpose_x_0"), val = bool(false)];
            bool var_7064_transpose_y_0 = const()[name = string("op_7064_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_6919_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7064_cast_fp16 = matmul(transpose_x = var_7064_transpose_x_0, transpose_y = var_7064_transpose_y_0, x = var_7063_cast_fp16, y = v_37_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7058_cast_fp16, y = var_7064_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7081_begin_0 = const()[name = string("op_7081_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7081_end_0 = const()[name = string("op_7081_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7081_end_mask_0 = const()[name = string("op_7081_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7081_cast_fp16 = slice_by_index(begin = var_7081_begin_0, end = var_7081_end_0, end_mask = var_7081_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7081_cast_fp16")];
            tensor<int32, [4]> var_7096_begin_0 = const()[name = string("op_7096_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7096_end_0 = const()[name = string("op_7096_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7096_end_mask_0 = const()[name = string("op_7096_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7096_cast_fp16 = slice_by_index(begin = var_7096_begin_0, end = var_7096_end_0, end_mask = var_7096_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7096_cast_fp16")];
            int32 var_7098 = const()[name = string("op_7098"), val = int32(1)];
            bool mixed_75_interleave_0 = const()[name = string("mixed_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_75_cast_fp16 = concat(axis = var_7098, interleave = mixed_75_interleave_0, values = (var_7081_cast_fp16, var_7096_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7102_perm_0 = const()[name = string("op_7102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7106 = const()[name = string("op_7106"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7102_cast_fp16 = transpose(perm = var_7102_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7106, x = var_7102_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599115840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601213056))))[name = string("model_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = linear_129_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_7117_promoted_to_fp16 = const()[name = string("op_7117_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7119_cast_fp16 = pow(x = x_227_cast_fp16, y = var_7117_promoted_to_fp16)[name = string("op_7119_cast_fp16")];
            tensor<int32, [1]> var_7121_axes_0 = const()[name = string("op_7121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7121_keep_dims_0 = const()[name = string("op_7121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7121_cast_fp16 = reduce_mean(axes = var_7121_axes_0, keep_dims = var_7121_keep_dims_0, x = var_7119_cast_fp16)[name = string("op_7121_cast_fp16")];
            fp16 var_7122_to_fp16 = const()[name = string("op_7122_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7123_cast_fp16 = add(x = var_7121_cast_fp16, y = var_7122_to_fp16)[name = string("op_7123_cast_fp16")];
            fp32 var_7124_epsilon_0 = const()[name = string("op_7124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7124_cast_fp16 = rsqrt(epsilon = var_7124_epsilon_0, x = var_7123_cast_fp16)[name = string("op_7124_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7125_cast_fp16 = mul(x = x_227_cast_fp16, y = var_7124_cast_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601215168)))];
            tensor<fp16, [1, 16, 1024]> input_185_cast_fp16 = mul(x = var_7125_cast_fp16, y = model_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601217280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604363072))))[name = string("model_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7130_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7130_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604369280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607515072))))[name = string("model_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_189_cast_fp16 = mul(x = var_7130_cast_fp16, y = linear_131_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607521280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610667072))))[name = string("model_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_132_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_229_cast_fp16 = add(x = x_227_cast_fp16, y = linear_132_cast_fp16)[name = string("x_229_cast_fp16")];
            fp16 var_7144_promoted_to_fp16 = const()[name = string("op_7144_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7146_cast_fp16 = pow(x = x_229_cast_fp16, y = var_7144_promoted_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<int32, [1]> var_7148_axes_0 = const()[name = string("op_7148_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7148_keep_dims_0 = const()[name = string("op_7148_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7148_cast_fp16 = reduce_mean(axes = var_7148_axes_0, keep_dims = var_7148_keep_dims_0, x = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp16 var_7149_to_fp16 = const()[name = string("op_7149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7150_cast_fp16 = add(x = var_7148_cast_fp16, y = var_7149_to_fp16)[name = string("op_7150_cast_fp16")];
            fp32 var_7151_epsilon_0 = const()[name = string("op_7151_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7151_cast_fp16 = rsqrt(epsilon = var_7151_epsilon_0, x = var_7150_cast_fp16)[name = string("op_7151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7152_cast_fp16 = mul(x = x_229_cast_fp16, y = var_7151_cast_fp16)[name = string("op_7152_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610669184)))];
            tensor<fp16, [1, 16, 1024]> input_191_cast_fp16 = mul(x = var_7152_cast_fp16, y = model_layers_19_input_layernorm_weight_to_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610671296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612768512))))[name = string("model_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_7161 = const()[name = string("op_7161"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7162_cast_fp16 = reshape(shape = var_7161, x = linear_133_cast_fp16)[name = string("op_7162_cast_fp16")];
            tensor<int32, [4]> x_231_perm_0 = const()[name = string("x_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7170_promoted_to_fp16 = const()[name = string("op_7170_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_231_cast_fp16 = transpose(perm = x_231_perm_0, x = var_7162_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7172_cast_fp16 = pow(x = x_231_cast_fp16, y = var_7170_promoted_to_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [1]> var_7174_axes_0 = const()[name = string("op_7174_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7174_keep_dims_0 = const()[name = string("op_7174_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7174_cast_fp16 = reduce_mean(axes = var_7174_axes_0, keep_dims = var_7174_keep_dims_0, x = var_7172_cast_fp16)[name = string("op_7174_cast_fp16")];
            fp16 var_7175_to_fp16 = const()[name = string("op_7175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7176_cast_fp16 = add(x = var_7174_cast_fp16, y = var_7175_to_fp16)[name = string("op_7176_cast_fp16")];
            fp32 var_7177_epsilon_0 = const()[name = string("op_7177_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7177_cast_fp16 = rsqrt(epsilon = var_7177_epsilon_0, x = var_7176_cast_fp16)[name = string("op_7177_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7178_cast_fp16 = mul(x = x_231_cast_fp16, y = var_7177_cast_fp16)[name = string("op_7178_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772672)))];
            tensor<fp16, [1, 16, 16, 128]> x_235_cast_fp16 = mul(x = var_7178_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613821632))))[name = string("model_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_7187 = const()[name = string("op_7187"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7188_cast_fp16 = reshape(shape = var_7187, x = linear_134_cast_fp16)[name = string("op_7188_cast_fp16")];
            tensor<int32, [4]> x_233_perm_0 = const()[name = string("x_233_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7196_promoted_to_fp16 = const()[name = string("op_7196_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_233_cast_fp16 = transpose(perm = x_233_perm_0, x = var_7188_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7198_cast_fp16 = pow(x = x_233_cast_fp16, y = var_7196_promoted_to_fp16)[name = string("op_7198_cast_fp16")];
            tensor<int32, [1]> var_7200_axes_0 = const()[name = string("op_7200_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7200_keep_dims_0 = const()[name = string("op_7200_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7200_cast_fp16 = reduce_mean(axes = var_7200_axes_0, keep_dims = var_7200_keep_dims_0, x = var_7198_cast_fp16)[name = string("op_7200_cast_fp16")];
            fp16 var_7201_to_fp16 = const()[name = string("op_7201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7202_cast_fp16 = add(x = var_7200_cast_fp16, y = var_7201_to_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7203_cast_fp16 = rsqrt(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7204_cast_fp16 = mul(x = x_233_cast_fp16, y = var_7203_cast_fp16)[name = string("op_7204_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613823744)))];
            tensor<fp16, [1, 8, 16, 128]> x_237_cast_fp16 = mul(x = var_7204_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613824064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614872704))))[name = string("model_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_7213 = const()[name = string("op_7213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7214_cast_fp16 = reshape(shape = var_7213, x = linear_135_cast_fp16)[name = string("op_7214_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_77_begin_0 = const()[name = string("first_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_77_end_0 = const()[name = string("first_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_77_end_mask_0 = const()[name = string("first_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_77_cast_fp16 = slice_by_index(begin = first_77_begin_0, end = first_77_end_0, end_mask = first_77_end_mask_0, x = x_235_cast_fp16)[name = string("first_77_cast_fp16")];
            tensor<int32, [4]> second_77_begin_0 = const()[name = string("second_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_77_end_0 = const()[name = string("second_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_77_end_mask_0 = const()[name = string("second_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_77_cast_fp16 = slice_by_index(begin = second_77_begin_0, end = second_77_end_0, end_mask = second_77_end_mask_0, x = x_235_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7228_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7229_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7231_cast_fp16 = sub(x = var_7228_cast_fp16, y = var_7229_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7232_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7233_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7235_cast_fp16 = add(x = var_7232_cast_fp16, y = var_7233_cast_fp16)[name = string("op_7235_cast_fp16")];
            int32 var_7237 = const()[name = string("op_7237"), val = int32(-1)];
            bool q_39_interleave_0 = const()[name = string("q_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_39_cast_fp16 = concat(axis = var_7237, interleave = q_39_interleave_0, values = (var_7231_cast_fp16, var_7235_cast_fp16))[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> first_79_begin_0 = const()[name = string("first_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_79_end_0 = const()[name = string("first_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_79_end_mask_0 = const()[name = string("first_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_79_cast_fp16 = slice_by_index(begin = first_79_begin_0, end = first_79_end_0, end_mask = first_79_end_mask_0, x = x_237_cast_fp16)[name = string("first_79_cast_fp16")];
            tensor<int32, [4]> second_79_begin_0 = const()[name = string("second_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_79_end_0 = const()[name = string("second_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_79_end_mask_0 = const()[name = string("second_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_79_cast_fp16 = slice_by_index(begin = second_79_begin_0, end = second_79_end_0, end_mask = second_79_end_mask_0, x = x_237_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7249_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7250_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7252_cast_fp16 = sub(x = var_7249_cast_fp16, y = var_7250_cast_fp16)[name = string("op_7252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7253_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7254_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7256_cast_fp16 = add(x = var_7253_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7256_cast_fp16")];
            int32 var_7258 = const()[name = string("op_7258"), val = int32(-1)];
            bool k_39_interleave_0 = const()[name = string("k_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_39_cast_fp16 = concat(axis = var_7258, interleave = k_39_interleave_0, values = (var_7252_cast_fp16, var_7256_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7269_begin_0 = const()[name = string("op_7269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7269_end_0 = const()[name = string("op_7269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7269_end_mask_0 = const()[name = string("op_7269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7269_stride_0 = const()[name = string("op_7269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7269_cast_fp16 = slice_by_index(begin = var_7269_begin_0, end = var_7269_end_0, end_mask = var_7269_end_mask_0, stride = var_7269_stride_0, x = q_39_cast_fp16)[name = string("op_7269_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7279_stride_0 = const()[name = string("op_7279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, stride = var_7279_stride_0, x = q_39_cast_fp16)[name = string("op_7279_cast_fp16")];
            int32 var_7281 = const()[name = string("op_7281"), val = int32(2)];
            bool stacked_39_interleave_0 = const()[name = string("stacked_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_39_cast_fp16 = concat(axis = var_7281, interleave = stacked_39_interleave_0, values = (var_7269_cast_fp16, var_7279_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7287_begin_0 = const()[name = string("op_7287_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7287_end_0 = const()[name = string("op_7287_end_0"), val = tensor<int32, [4]>([20, 8, 4112, 128])];
            tensor<bool, [4]> var_7287_end_mask_0 = const()[name = string("op_7287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7287_cast_fp16 = slice_by_index(begin = var_7287_begin_0, end = var_7287_end_0, end_mask = var_7287_end_mask_0, x = read_state_0)[name = string("op_7287_cast_fp16")];
            tensor<int32, [4]> var_7297_begin_0 = const()[name = string("op_7297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7297_end_0 = const()[name = string("op_7297_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7297_end_mask_0 = const()[name = string("op_7297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7297_cast_fp16 = slice_by_index(begin = var_7297_begin_0, end = var_7297_end_0, end_mask = var_7297_end_mask_0, x = var_7287_cast_fp16)[name = string("op_7297_cast_fp16")];
            bool var_7306_transpose_x_1 = const()[name = string("op_7306_transpose_x_1"), val = bool(false)];
            bool var_7306_transpose_y_1 = const()[name = string("op_7306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7306_cast_fp16 = matmul(transpose_x = var_7306_transpose_x_1, transpose_y = var_7306_transpose_y_1, x = var_7297_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7306_cast_fp16")];
            tensor<int32, [4]> var_7309_perm_0 = const()[name = string("op_7309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7310_to_fp16 = const()[name = string("op_7310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7309_cast_fp16 = transpose(perm = var_7309_perm_0, x = var_7306_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 4096]> var_7311_cast_fp16 = mul(x = var_7309_cast_fp16, y = var_7310_to_fp16)[name = string("op_7311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_39_cast_fp16 = add(x = var_7311_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7317_transpose_x_1 = const()[name = string("op_7317_transpose_x_1"), val = bool(false)];
            bool var_7317_transpose_y_1 = const()[name = string("op_7317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7317_cast_fp16 = matmul(transpose_x = var_7317_transpose_x_1, transpose_y = var_7317_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7317_cast_fp16")];
            fp16 var_7318_to_fp16 = const()[name = string("op_7318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7319_cast_fp16 = mul(x = var_7317_cast_fp16, y = var_7318_to_fp16)[name = string("op_7319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7323 = const()[name = string("op_7323"), val = int32(-1)];
            bool var_7324_interleave_0 = const()[name = string("op_7324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7324_cast_fp16 = concat(axis = var_7323, interleave = var_7324_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7324_cast_fp16")];
            int32 var_7325 = const()[name = string("op_7325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_39_cast_fp16 = softmax(axis = var_7325, x = var_7324_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7332_begin_0 = const()[name = string("op_7332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7332_end_0 = const()[name = string("op_7332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7332_end_mask_0 = const()[name = string("op_7332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7332_cast_fp16 = slice_by_index(begin = var_7332_begin_0, end = var_7332_end_0, end_mask = var_7332_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7332_cast_fp16")];
            tensor<int32, [4]> var_7337_begin_0 = const()[name = string("op_7337_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7337_end_0 = const()[name = string("op_7337_end_0"), val = tensor<int32, [4]>([20, 8, 4112, 128])];
            tensor<bool, [4]> var_7337_end_mask_0 = const()[name = string("op_7337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7337_cast_fp16 = slice_by_index(begin = var_7337_begin_0, end = var_7337_end_0, end_mask = var_7337_end_mask_0, x = read_state_1)[name = string("op_7337_cast_fp16")];
            tensor<int32, [4]> var_7347_begin_0 = const()[name = string("op_7347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7347_end_0 = const()[name = string("op_7347_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7347_end_mask_0 = const()[name = string("op_7347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7347_cast_fp16 = slice_by_index(begin = var_7347_begin_0, end = var_7347_end_0, end_mask = var_7347_end_mask_0, x = var_7337_cast_fp16)[name = string("op_7347_cast_fp16")];
            bool var_7353_transpose_x_0 = const()[name = string("op_7353_transpose_x_0"), val = bool(false)];
            bool var_7353_transpose_y_0 = const()[name = string("op_7353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7353_cast_fp16 = matmul(transpose_x = var_7353_transpose_x_0, transpose_y = var_7353_transpose_y_0, x = var_7332_cast_fp16, y = var_7347_cast_fp16)[name = string("op_7353_cast_fp16")];
            tensor<int32, [4]> var_7358_begin_0 = const()[name = string("op_7358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7358_end_0 = const()[name = string("op_7358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7358_end_mask_0 = const()[name = string("op_7358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7358_cast_fp16 = slice_by_index(begin = var_7358_begin_0, end = var_7358_end_0, end_mask = var_7358_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7358_cast_fp16")];
            bool var_7359_transpose_x_0 = const()[name = string("op_7359_transpose_x_0"), val = bool(false)];
            bool var_7359_transpose_y_0 = const()[name = string("op_7359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7214_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7359_cast_fp16 = matmul(transpose_x = var_7359_transpose_x_0, transpose_y = var_7359_transpose_y_0, x = var_7358_cast_fp16, y = v_39_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7353_cast_fp16, y = var_7359_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7376_begin_0 = const()[name = string("op_7376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7376_end_0 = const()[name = string("op_7376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7376_end_mask_0 = const()[name = string("op_7376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7376_cast_fp16 = slice_by_index(begin = var_7376_begin_0, end = var_7376_end_0, end_mask = var_7376_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7376_cast_fp16")];
            tensor<int32, [4]> var_7391_begin_0 = const()[name = string("op_7391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7391_end_0 = const()[name = string("op_7391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7391_end_mask_0 = const()[name = string("op_7391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7391_cast_fp16")];
            int32 var_7393 = const()[name = string("op_7393"), val = int32(1)];
            bool mixed_79_interleave_0 = const()[name = string("mixed_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_79_cast_fp16 = concat(axis = var_7393, interleave = mixed_79_interleave_0, values = (var_7376_cast_fp16, var_7391_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7397_perm_0 = const()[name = string("op_7397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7401 = const()[name = string("op_7401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7397_cast_fp16 = transpose(perm = var_7397_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7401, x = var_7397_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614874816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616972032))))[name = string("model_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = add(x = x_229_cast_fp16, y = linear_136_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_7412_promoted_to_fp16 = const()[name = string("op_7412_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7414_cast_fp16 = pow(x = x_239_cast_fp16, y = var_7412_promoted_to_fp16)[name = string("op_7414_cast_fp16")];
            tensor<int32, [1]> var_7416_axes_0 = const()[name = string("op_7416_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7416_keep_dims_0 = const()[name = string("op_7416_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7416_cast_fp16 = reduce_mean(axes = var_7416_axes_0, keep_dims = var_7416_keep_dims_0, x = var_7414_cast_fp16)[name = string("op_7416_cast_fp16")];
            fp16 var_7417_to_fp16 = const()[name = string("op_7417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7418_cast_fp16 = add(x = var_7416_cast_fp16, y = var_7417_to_fp16)[name = string("op_7418_cast_fp16")];
            fp32 var_7419_epsilon_0 = const()[name = string("op_7419_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7419_cast_fp16 = rsqrt(epsilon = var_7419_epsilon_0, x = var_7418_cast_fp16)[name = string("op_7419_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7420_cast_fp16 = mul(x = x_239_cast_fp16, y = var_7419_cast_fp16)[name = string("op_7420_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616974144)))];
            tensor<fp16, [1, 16, 1024]> input_195_cast_fp16 = mul(x = var_7420_cast_fp16, y = model_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616976256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620122048))))[name = string("model_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7425_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7425_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620128256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623274048))))[name = string("model_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_199_cast_fp16 = mul(x = var_7425_cast_fp16, y = linear_138_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623280256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626426048))))[name = string("model_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_241_cast_fp16 = add(x = x_239_cast_fp16, y = linear_139_cast_fp16)[name = string("x_241_cast_fp16")];
            fp16 var_7439_promoted_to_fp16 = const()[name = string("op_7439_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7441_cast_fp16 = pow(x = x_241_cast_fp16, y = var_7439_promoted_to_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7443_axes_0 = const()[name = string("op_7443_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7443_keep_dims_0 = const()[name = string("op_7443_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7443_cast_fp16 = reduce_mean(axes = var_7443_axes_0, keep_dims = var_7443_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7443_cast_fp16")];
            fp16 var_7444_to_fp16 = const()[name = string("op_7444_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7445_cast_fp16 = add(x = var_7443_cast_fp16, y = var_7444_to_fp16)[name = string("op_7445_cast_fp16")];
            fp32 var_7446_epsilon_0 = const()[name = string("op_7446_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7446_cast_fp16 = rsqrt(epsilon = var_7446_epsilon_0, x = var_7445_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7447_cast_fp16 = mul(x = x_241_cast_fp16, y = var_7446_cast_fp16)[name = string("op_7447_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626428160)))];
            tensor<fp16, [1, 16, 1024]> input_201_cast_fp16 = mul(x = var_7447_cast_fp16, y = model_layers_20_input_layernorm_weight_to_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626430272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628527488))))[name = string("model_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_7456 = const()[name = string("op_7456"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7457_cast_fp16 = reshape(shape = var_7456, x = linear_140_cast_fp16)[name = string("op_7457_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7465_promoted_to_fp16 = const()[name = string("op_7465_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_7457_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_7467_cast_fp16 = pow(x = x_243_cast_fp16, y = var_7465_promoted_to_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7469_axes_0 = const()[name = string("op_7469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7469_keep_dims_0 = const()[name = string("op_7469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7469_cast_fp16 = reduce_mean(axes = var_7469_axes_0, keep_dims = var_7469_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7469_cast_fp16")];
            fp16 var_7470_to_fp16 = const()[name = string("op_7470_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7471_cast_fp16 = add(x = var_7469_cast_fp16, y = var_7470_to_fp16)[name = string("op_7471_cast_fp16")];
            fp32 var_7472_epsilon_0 = const()[name = string("op_7472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7472_cast_fp16 = rsqrt(epsilon = var_7472_epsilon_0, x = var_7471_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7473_cast_fp16 = mul(x = x_243_cast_fp16, y = var_7472_cast_fp16)[name = string("op_7473_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531648)))];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = mul(x = var_7473_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_247_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629580608))))[name = string("model_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_7482 = const()[name = string("op_7482"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7483_cast_fp16 = reshape(shape = var_7482, x = linear_141_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7491_promoted_to_fp16 = const()[name = string("op_7491_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_7483_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_7493_cast_fp16 = pow(x = x_245_cast_fp16, y = var_7491_promoted_to_fp16)[name = string("op_7493_cast_fp16")];
            tensor<int32, [1]> var_7495_axes_0 = const()[name = string("op_7495_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7495_keep_dims_0 = const()[name = string("op_7495_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7495_cast_fp16 = reduce_mean(axes = var_7495_axes_0, keep_dims = var_7495_keep_dims_0, x = var_7493_cast_fp16)[name = string("op_7495_cast_fp16")];
            fp16 var_7496_to_fp16 = const()[name = string("op_7496_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7497_cast_fp16 = add(x = var_7495_cast_fp16, y = var_7496_to_fp16)[name = string("op_7497_cast_fp16")];
            fp32 var_7498_epsilon_0 = const()[name = string("op_7498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7498_cast_fp16 = rsqrt(epsilon = var_7498_epsilon_0, x = var_7497_cast_fp16)[name = string("op_7498_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = mul(x = x_245_cast_fp16, y = var_7498_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629582720)))];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = mul(x = var_7499_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629583040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630631680))))[name = string("model_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_7508 = const()[name = string("op_7508"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7509_cast_fp16 = reshape(shape = var_7508, x = linear_142_cast_fp16)[name = string("op_7509_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_81_begin_0 = const()[name = string("first_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_81_end_0 = const()[name = string("first_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_81_end_mask_0 = const()[name = string("first_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_81_cast_fp16 = slice_by_index(begin = first_81_begin_0, end = first_81_end_0, end_mask = first_81_end_mask_0, x = x_247_cast_fp16)[name = string("first_81_cast_fp16")];
            tensor<int32, [4]> second_81_begin_0 = const()[name = string("second_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_81_end_0 = const()[name = string("second_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_81_end_mask_0 = const()[name = string("second_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_81_cast_fp16 = slice_by_index(begin = second_81_begin_0, end = second_81_end_0, end_mask = second_81_end_mask_0, x = x_247_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7523_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7524_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7526_cast_fp16 = sub(x = var_7523_cast_fp16, y = var_7524_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7527_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7527_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7528_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7528_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7530_cast_fp16 = add(x = var_7527_cast_fp16, y = var_7528_cast_fp16)[name = string("op_7530_cast_fp16")];
            int32 var_7532 = const()[name = string("op_7532"), val = int32(-1)];
            bool q_41_interleave_0 = const()[name = string("q_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_41_cast_fp16 = concat(axis = var_7532, interleave = q_41_interleave_0, values = (var_7526_cast_fp16, var_7530_cast_fp16))[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> first_83_begin_0 = const()[name = string("first_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_83_end_0 = const()[name = string("first_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_83_end_mask_0 = const()[name = string("first_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_83_cast_fp16 = slice_by_index(begin = first_83_begin_0, end = first_83_end_0, end_mask = first_83_end_mask_0, x = x_249_cast_fp16)[name = string("first_83_cast_fp16")];
            tensor<int32, [4]> second_83_begin_0 = const()[name = string("second_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_83_end_0 = const()[name = string("second_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_83_end_mask_0 = const()[name = string("second_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_83_cast_fp16 = slice_by_index(begin = second_83_begin_0, end = second_83_end_0, end_mask = second_83_end_mask_0, x = x_249_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7544_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7545_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7547_cast_fp16 = sub(x = var_7544_cast_fp16, y = var_7545_cast_fp16)[name = string("op_7547_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7548_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7549_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7551_cast_fp16 = add(x = var_7548_cast_fp16, y = var_7549_cast_fp16)[name = string("op_7551_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool k_41_interleave_0 = const()[name = string("k_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_41_cast_fp16 = concat(axis = var_7553, interleave = k_41_interleave_0, values = (var_7547_cast_fp16, var_7551_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_7564_begin_0 = const()[name = string("op_7564_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7564_end_0 = const()[name = string("op_7564_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7564_end_mask_0 = const()[name = string("op_7564_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7564_stride_0 = const()[name = string("op_7564_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7564_cast_fp16 = slice_by_index(begin = var_7564_begin_0, end = var_7564_end_0, end_mask = var_7564_end_mask_0, stride = var_7564_stride_0, x = q_41_cast_fp16)[name = string("op_7564_cast_fp16")];
            tensor<int32, [4]> var_7574_begin_0 = const()[name = string("op_7574_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7574_end_0 = const()[name = string("op_7574_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7574_end_mask_0 = const()[name = string("op_7574_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7574_stride_0 = const()[name = string("op_7574_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7574_cast_fp16 = slice_by_index(begin = var_7574_begin_0, end = var_7574_end_0, end_mask = var_7574_end_mask_0, stride = var_7574_stride_0, x = q_41_cast_fp16)[name = string("op_7574_cast_fp16")];
            int32 var_7576 = const()[name = string("op_7576"), val = int32(2)];
            bool stacked_41_interleave_0 = const()[name = string("stacked_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_41_cast_fp16 = concat(axis = var_7576, interleave = stacked_41_interleave_0, values = (var_7564_cast_fp16, var_7574_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_7582_begin_0 = const()[name = string("op_7582_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7582_end_0 = const()[name = string("op_7582_end_0"), val = tensor<int32, [4]>([21, 8, 4112, 128])];
            tensor<bool, [4]> var_7582_end_mask_0 = const()[name = string("op_7582_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7582_cast_fp16 = slice_by_index(begin = var_7582_begin_0, end = var_7582_end_0, end_mask = var_7582_end_mask_0, x = read_state_0)[name = string("op_7582_cast_fp16")];
            tensor<int32, [4]> var_7592_begin_0 = const()[name = string("op_7592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7592_end_0 = const()[name = string("op_7592_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7592_end_mask_0 = const()[name = string("op_7592_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7592_cast_fp16 = slice_by_index(begin = var_7592_begin_0, end = var_7592_end_0, end_mask = var_7592_end_mask_0, x = var_7582_cast_fp16)[name = string("op_7592_cast_fp16")];
            bool var_7601_transpose_x_1 = const()[name = string("op_7601_transpose_x_1"), val = bool(false)];
            bool var_7601_transpose_y_1 = const()[name = string("op_7601_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7601_cast_fp16 = matmul(transpose_x = var_7601_transpose_x_1, transpose_y = var_7601_transpose_y_1, x = var_7592_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_7601_cast_fp16")];
            tensor<int32, [4]> var_7604_perm_0 = const()[name = string("op_7604_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7605_to_fp16 = const()[name = string("op_7605_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7604_cast_fp16 = transpose(perm = var_7604_perm_0, x = var_7601_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 4096]> var_7606_cast_fp16 = mul(x = var_7604_cast_fp16, y = var_7605_to_fp16)[name = string("op_7606_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_41_cast_fp16 = add(x = var_7606_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_7612_transpose_x_1 = const()[name = string("op_7612_transpose_x_1"), val = bool(false)];
            bool var_7612_transpose_y_1 = const()[name = string("op_7612_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7612_cast_fp16 = matmul(transpose_x = var_7612_transpose_x_1, transpose_y = var_7612_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_7612_cast_fp16")];
            fp16 var_7613_to_fp16 = const()[name = string("op_7613_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7614_cast_fp16 = mul(x = var_7612_cast_fp16, y = var_7613_to_fp16)[name = string("op_7614_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_7614_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_7618 = const()[name = string("op_7618"), val = int32(-1)];
            bool var_7619_interleave_0 = const()[name = string("op_7619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7619_cast_fp16 = concat(axis = var_7618, interleave = var_7619_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_7619_cast_fp16")];
            int32 var_7620 = const()[name = string("op_7620"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_41_cast_fp16 = softmax(axis = var_7620, x = var_7619_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_7627_begin_0 = const()[name = string("op_7627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7627_end_0 = const()[name = string("op_7627_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7627_end_mask_0 = const()[name = string("op_7627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7627_cast_fp16 = slice_by_index(begin = var_7627_begin_0, end = var_7627_end_0, end_mask = var_7627_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7627_cast_fp16")];
            tensor<int32, [4]> var_7632_begin_0 = const()[name = string("op_7632_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7632_end_0 = const()[name = string("op_7632_end_0"), val = tensor<int32, [4]>([21, 8, 4112, 128])];
            tensor<bool, [4]> var_7632_end_mask_0 = const()[name = string("op_7632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7632_cast_fp16 = slice_by_index(begin = var_7632_begin_0, end = var_7632_end_0, end_mask = var_7632_end_mask_0, x = read_state_1)[name = string("op_7632_cast_fp16")];
            tensor<int32, [4]> var_7642_begin_0 = const()[name = string("op_7642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7642_end_0 = const()[name = string("op_7642_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7642_end_mask_0 = const()[name = string("op_7642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7642_cast_fp16 = slice_by_index(begin = var_7642_begin_0, end = var_7642_end_0, end_mask = var_7642_end_mask_0, x = var_7632_cast_fp16)[name = string("op_7642_cast_fp16")];
            bool var_7648_transpose_x_0 = const()[name = string("op_7648_transpose_x_0"), val = bool(false)];
            bool var_7648_transpose_y_0 = const()[name = string("op_7648_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7648_cast_fp16 = matmul(transpose_x = var_7648_transpose_x_0, transpose_y = var_7648_transpose_y_0, x = var_7627_cast_fp16, y = var_7642_cast_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [4]> var_7653_begin_0 = const()[name = string("op_7653_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7653_end_0 = const()[name = string("op_7653_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7653_end_mask_0 = const()[name = string("op_7653_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7653_cast_fp16 = slice_by_index(begin = var_7653_begin_0, end = var_7653_end_0, end_mask = var_7653_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7653_cast_fp16")];
            bool var_7654_transpose_x_0 = const()[name = string("op_7654_transpose_x_0"), val = bool(false)];
            bool var_7654_transpose_y_0 = const()[name = string("op_7654_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_7509_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_7654_cast_fp16 = matmul(transpose_x = var_7654_transpose_x_0, transpose_y = var_7654_transpose_y_0, x = var_7653_cast_fp16, y = v_41_cast_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_7648_cast_fp16, y = var_7654_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_7671_begin_0 = const()[name = string("op_7671_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7671_end_0 = const()[name = string("op_7671_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7671_end_mask_0 = const()[name = string("op_7671_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7671_cast_fp16 = slice_by_index(begin = var_7671_begin_0, end = var_7671_end_0, end_mask = var_7671_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<int32, [4]> var_7686_begin_0 = const()[name = string("op_7686_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7686_end_0 = const()[name = string("op_7686_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7686_end_mask_0 = const()[name = string("op_7686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7686_cast_fp16 = slice_by_index(begin = var_7686_begin_0, end = var_7686_end_0, end_mask = var_7686_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7686_cast_fp16")];
            int32 var_7688 = const()[name = string("op_7688"), val = int32(1)];
            bool mixed_83_interleave_0 = const()[name = string("mixed_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_83_cast_fp16 = concat(axis = var_7688, interleave = mixed_83_interleave_0, values = (var_7671_cast_fp16, var_7686_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_7692_perm_0 = const()[name = string("op_7692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7696 = const()[name = string("op_7696"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7692_cast_fp16 = transpose(perm = var_7692_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_7696, x = var_7692_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630633792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632731008))))[name = string("model_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_251_cast_fp16 = add(x = x_241_cast_fp16, y = linear_143_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7709_cast_fp16 = pow(x = x_251_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7709_cast_fp16")];
            tensor<int32, [1]> var_7711_axes_0 = const()[name = string("op_7711_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7711_keep_dims_0 = const()[name = string("op_7711_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7711_cast_fp16 = reduce_mean(axes = var_7711_axes_0, keep_dims = var_7711_keep_dims_0, x = var_7709_cast_fp16)[name = string("op_7711_cast_fp16")];
            fp16 var_7712_to_fp16 = const()[name = string("op_7712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7713_cast_fp16 = add(x = var_7711_cast_fp16, y = var_7712_to_fp16)[name = string("op_7713_cast_fp16")];
            fp32 var_7714_epsilon_0 = const()[name = string("op_7714_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7714_cast_fp16 = rsqrt(epsilon = var_7714_epsilon_0, x = var_7713_cast_fp16)[name = string("op_7714_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7715_cast_fp16 = mul(x = x_251_cast_fp16, y = var_7714_cast_fp16)[name = string("op_7715_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632733120)))];
            tensor<fp16, [1, 16, 1024]> input_205_cast_fp16 = mul(x = var_7715_cast_fp16, y = model_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632735232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635881024))))[name = string("model_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7720_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_7720_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635887232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639033024))))[name = string("model_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_209_cast_fp16 = mul(x = var_7720_cast_fp16, y = linear_145_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639039232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642185024))))[name = string("model_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_146_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_253_cast_fp16 = add(x = x_251_cast_fp16, y = linear_146_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_7734_promoted_to_fp16 = const()[name = string("op_7734_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7736_cast_fp16 = pow(x = x_253_cast_fp16, y = var_7734_promoted_to_fp16)[name = string("op_7736_cast_fp16")];
            tensor<int32, [1]> var_7738_axes_0 = const()[name = string("op_7738_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7738_keep_dims_0 = const()[name = string("op_7738_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7738_cast_fp16 = reduce_mean(axes = var_7738_axes_0, keep_dims = var_7738_keep_dims_0, x = var_7736_cast_fp16)[name = string("op_7738_cast_fp16")];
            fp16 var_7739_to_fp16 = const()[name = string("op_7739_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7740_cast_fp16 = add(x = var_7738_cast_fp16, y = var_7739_to_fp16)[name = string("op_7740_cast_fp16")];
            fp32 var_7741_epsilon_0 = const()[name = string("op_7741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7741_cast_fp16 = rsqrt(epsilon = var_7741_epsilon_0, x = var_7740_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7742_cast_fp16 = mul(x = x_253_cast_fp16, y = var_7741_cast_fp16)[name = string("op_7742_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642187136)))];
            tensor<fp16, [1, 16, 1024]> input_211_cast_fp16 = mul(x = var_7742_cast_fp16, y = model_layers_21_input_layernorm_weight_to_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642189248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644286464))))[name = string("model_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7751 = const()[name = string("op_7751"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7752_cast_fp16 = reshape(shape = var_7751, x = linear_147_cast_fp16)[name = string("op_7752_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7760_promoted_to_fp16 = const()[name = string("op_7760_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_7752_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_7762_cast_fp16 = pow(x = x_255_cast_fp16, y = var_7760_promoted_to_fp16)[name = string("op_7762_cast_fp16")];
            tensor<int32, [1]> var_7764_axes_0 = const()[name = string("op_7764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7764_keep_dims_0 = const()[name = string("op_7764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7764_cast_fp16 = reduce_mean(axes = var_7764_axes_0, keep_dims = var_7764_keep_dims_0, x = var_7762_cast_fp16)[name = string("op_7764_cast_fp16")];
            fp16 var_7765_to_fp16 = const()[name = string("op_7765_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7766_cast_fp16 = add(x = var_7764_cast_fp16, y = var_7765_to_fp16)[name = string("op_7766_cast_fp16")];
            fp32 var_7767_epsilon_0 = const()[name = string("op_7767_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7767_cast_fp16 = rsqrt(epsilon = var_7767_epsilon_0, x = var_7766_cast_fp16)[name = string("op_7767_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7768_cast_fp16 = mul(x = x_255_cast_fp16, y = var_7767_cast_fp16)[name = string("op_7768_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290624)))];
            tensor<fp16, [1, 16, 16, 128]> x_259_cast_fp16 = mul(x = var_7768_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645339584))))[name = string("model_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_7777 = const()[name = string("op_7777"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7778_cast_fp16 = reshape(shape = var_7777, x = linear_148_cast_fp16)[name = string("op_7778_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7786_promoted_to_fp16 = const()[name = string("op_7786_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_7778_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_7788_cast_fp16 = pow(x = x_257_cast_fp16, y = var_7786_promoted_to_fp16)[name = string("op_7788_cast_fp16")];
            tensor<int32, [1]> var_7790_axes_0 = const()[name = string("op_7790_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7790_keep_dims_0 = const()[name = string("op_7790_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7790_cast_fp16 = reduce_mean(axes = var_7790_axes_0, keep_dims = var_7790_keep_dims_0, x = var_7788_cast_fp16)[name = string("op_7790_cast_fp16")];
            fp16 var_7791_to_fp16 = const()[name = string("op_7791_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7792_cast_fp16 = add(x = var_7790_cast_fp16, y = var_7791_to_fp16)[name = string("op_7792_cast_fp16")];
            fp32 var_7793_epsilon_0 = const()[name = string("op_7793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7793_cast_fp16 = rsqrt(epsilon = var_7793_epsilon_0, x = var_7792_cast_fp16)[name = string("op_7793_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7794_cast_fp16 = mul(x = x_257_cast_fp16, y = var_7793_cast_fp16)[name = string("op_7794_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645341696)))];
            tensor<fp16, [1, 8, 16, 128]> x_261_cast_fp16 = mul(x = var_7794_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645342016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646390656))))[name = string("model_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_7803 = const()[name = string("op_7803"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7804_cast_fp16 = reshape(shape = var_7803, x = linear_149_cast_fp16)[name = string("op_7804_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_85_begin_0 = const()[name = string("first_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_85_end_0 = const()[name = string("first_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_85_end_mask_0 = const()[name = string("first_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_85_cast_fp16 = slice_by_index(begin = first_85_begin_0, end = first_85_end_0, end_mask = first_85_end_mask_0, x = x_259_cast_fp16)[name = string("first_85_cast_fp16")];
            tensor<int32, [4]> second_85_begin_0 = const()[name = string("second_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_85_end_0 = const()[name = string("second_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_85_end_mask_0 = const()[name = string("second_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_85_cast_fp16 = slice_by_index(begin = second_85_begin_0, end = second_85_end_0, end_mask = second_85_end_mask_0, x = x_259_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7818_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7818_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7819_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7819_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7821_cast_fp16 = sub(x = var_7818_cast_fp16, y = var_7819_cast_fp16)[name = string("op_7821_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7822_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7823_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7823_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7825_cast_fp16 = add(x = var_7822_cast_fp16, y = var_7823_cast_fp16)[name = string("op_7825_cast_fp16")];
            int32 var_7827 = const()[name = string("op_7827"), val = int32(-1)];
            bool q_43_interleave_0 = const()[name = string("q_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_43_cast_fp16 = concat(axis = var_7827, interleave = q_43_interleave_0, values = (var_7821_cast_fp16, var_7825_cast_fp16))[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> first_87_begin_0 = const()[name = string("first_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_87_end_0 = const()[name = string("first_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_87_end_mask_0 = const()[name = string("first_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_87_cast_fp16 = slice_by_index(begin = first_87_begin_0, end = first_87_end_0, end_mask = first_87_end_mask_0, x = x_261_cast_fp16)[name = string("first_87_cast_fp16")];
            tensor<int32, [4]> second_87_begin_0 = const()[name = string("second_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_87_end_0 = const()[name = string("second_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_87_end_mask_0 = const()[name = string("second_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_87_cast_fp16 = slice_by_index(begin = second_87_begin_0, end = second_87_end_0, end_mask = second_87_end_mask_0, x = x_261_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7839_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7839_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7840_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7840_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7842_cast_fp16 = sub(x = var_7839_cast_fp16, y = var_7840_cast_fp16)[name = string("op_7842_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7843_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7844_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7846_cast_fp16 = add(x = var_7843_cast_fp16, y = var_7844_cast_fp16)[name = string("op_7846_cast_fp16")];
            int32 var_7848 = const()[name = string("op_7848"), val = int32(-1)];
            bool k_43_interleave_0 = const()[name = string("k_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_43_cast_fp16 = concat(axis = var_7848, interleave = k_43_interleave_0, values = (var_7842_cast_fp16, var_7846_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_7859_begin_0 = const()[name = string("op_7859_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7859_end_0 = const()[name = string("op_7859_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7859_end_mask_0 = const()[name = string("op_7859_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7859_stride_0 = const()[name = string("op_7859_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7859_cast_fp16 = slice_by_index(begin = var_7859_begin_0, end = var_7859_end_0, end_mask = var_7859_end_mask_0, stride = var_7859_stride_0, x = q_43_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7869_begin_0 = const()[name = string("op_7869_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7869_end_0 = const()[name = string("op_7869_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7869_end_mask_0 = const()[name = string("op_7869_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7869_stride_0 = const()[name = string("op_7869_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7869_cast_fp16 = slice_by_index(begin = var_7869_begin_0, end = var_7869_end_0, end_mask = var_7869_end_mask_0, stride = var_7869_stride_0, x = q_43_cast_fp16)[name = string("op_7869_cast_fp16")];
            int32 var_7871 = const()[name = string("op_7871"), val = int32(2)];
            bool stacked_43_interleave_0 = const()[name = string("stacked_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_43_cast_fp16 = concat(axis = var_7871, interleave = stacked_43_interleave_0, values = (var_7859_cast_fp16, var_7869_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_7877_begin_0 = const()[name = string("op_7877_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7877_end_0 = const()[name = string("op_7877_end_0"), val = tensor<int32, [4]>([22, 8, 4112, 128])];
            tensor<bool, [4]> var_7877_end_mask_0 = const()[name = string("op_7877_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7877_cast_fp16 = slice_by_index(begin = var_7877_begin_0, end = var_7877_end_0, end_mask = var_7877_end_mask_0, x = read_state_0)[name = string("op_7877_cast_fp16")];
            tensor<int32, [4]> var_7887_begin_0 = const()[name = string("op_7887_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7887_end_0 = const()[name = string("op_7887_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7887_end_mask_0 = const()[name = string("op_7887_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7887_cast_fp16 = slice_by_index(begin = var_7887_begin_0, end = var_7887_end_0, end_mask = var_7887_end_mask_0, x = var_7877_cast_fp16)[name = string("op_7887_cast_fp16")];
            bool var_7896_transpose_x_1 = const()[name = string("op_7896_transpose_x_1"), val = bool(false)];
            bool var_7896_transpose_y_1 = const()[name = string("op_7896_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_7896_cast_fp16 = matmul(transpose_x = var_7896_transpose_x_1, transpose_y = var_7896_transpose_y_1, x = var_7887_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_7896_cast_fp16")];
            tensor<int32, [4]> var_7899_perm_0 = const()[name = string("op_7899_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7900_to_fp16 = const()[name = string("op_7900_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_7899_cast_fp16 = transpose(perm = var_7899_perm_0, x = var_7896_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 4096]> var_7901_cast_fp16 = mul(x = var_7899_cast_fp16, y = var_7900_to_fp16)[name = string("op_7901_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_43_cast_fp16 = add(x = var_7901_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_7907_transpose_x_1 = const()[name = string("op_7907_transpose_x_1"), val = bool(false)];
            bool var_7907_transpose_y_1 = const()[name = string("op_7907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7907_cast_fp16 = matmul(transpose_x = var_7907_transpose_x_1, transpose_y = var_7907_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_7907_cast_fp16")];
            fp16 var_7908_to_fp16 = const()[name = string("op_7908_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7909_cast_fp16 = mul(x = var_7907_cast_fp16, y = var_7908_to_fp16)[name = string("op_7909_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_7909_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_7913 = const()[name = string("op_7913"), val = int32(-1)];
            bool var_7914_interleave_0 = const()[name = string("op_7914_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_7914_cast_fp16 = concat(axis = var_7913, interleave = var_7914_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_7914_cast_fp16")];
            int32 var_7915 = const()[name = string("op_7915"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_43_cast_fp16 = softmax(axis = var_7915, x = var_7914_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_7922_begin_0 = const()[name = string("op_7922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7922_end_0 = const()[name = string("op_7922_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_7922_end_mask_0 = const()[name = string("op_7922_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([22, 8, 4112, 128])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = read_state_1)[name = string("op_7927_cast_fp16")];
            tensor<int32, [4]> var_7937_begin_0 = const()[name = string("op_7937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7937_end_0 = const()[name = string("op_7937_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_7937_end_mask_0 = const()[name = string("op_7937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_7937_cast_fp16 = slice_by_index(begin = var_7937_begin_0, end = var_7937_end_0, end_mask = var_7937_end_mask_0, x = var_7927_cast_fp16)[name = string("op_7937_cast_fp16")];
            bool var_7943_transpose_x_0 = const()[name = string("op_7943_transpose_x_0"), val = bool(false)];
            bool var_7943_transpose_y_0 = const()[name = string("op_7943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7943_cast_fp16 = matmul(transpose_x = var_7943_transpose_x_0, transpose_y = var_7943_transpose_y_0, x = var_7922_cast_fp16, y = var_7937_cast_fp16)[name = string("op_7943_cast_fp16")];
            tensor<int32, [4]> var_7948_begin_0 = const()[name = string("op_7948_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_7948_end_0 = const()[name = string("op_7948_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_7948_end_mask_0 = const()[name = string("op_7948_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7948_cast_fp16 = slice_by_index(begin = var_7948_begin_0, end = var_7948_end_0, end_mask = var_7948_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7948_cast_fp16")];
            bool var_7949_transpose_x_0 = const()[name = string("op_7949_transpose_x_0"), val = bool(false)];
            bool var_7949_transpose_y_0 = const()[name = string("op_7949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_7804_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_7949_cast_fp16 = matmul(transpose_x = var_7949_transpose_x_0, transpose_y = var_7949_transpose_y_0, x = var_7948_cast_fp16, y = v_43_cast_fp16)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_7943_cast_fp16, y = var_7949_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_7966_begin_0 = const()[name = string("op_7966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7966_end_0 = const()[name = string("op_7966_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7966_end_mask_0 = const()[name = string("op_7966_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7966_cast_fp16 = slice_by_index(begin = var_7966_begin_0, end = var_7966_end_0, end_mask = var_7966_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7966_cast_fp16")];
            tensor<int32, [4]> var_7981_begin_0 = const()[name = string("op_7981_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7981_end_0 = const()[name = string("op_7981_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7981_end_mask_0 = const()[name = string("op_7981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7981_cast_fp16 = slice_by_index(begin = var_7981_begin_0, end = var_7981_end_0, end_mask = var_7981_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7981_cast_fp16")];
            int32 var_7983 = const()[name = string("op_7983"), val = int32(1)];
            bool mixed_87_interleave_0 = const()[name = string("mixed_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_87_cast_fp16 = concat(axis = var_7983, interleave = mixed_87_interleave_0, values = (var_7966_cast_fp16, var_7981_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_7987_perm_0 = const()[name = string("op_7987_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7991 = const()[name = string("op_7991"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7987_cast_fp16 = transpose(perm = var_7987_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_7991, x = var_7987_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646392768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648489984))))[name = string("model_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_150_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = linear_150_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_8002_promoted_to_fp16 = const()[name = string("op_8002_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8004_cast_fp16 = pow(x = x_263_cast_fp16, y = var_8002_promoted_to_fp16)[name = string("op_8004_cast_fp16")];
            tensor<int32, [1]> var_8006_axes_0 = const()[name = string("op_8006_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8006_keep_dims_0 = const()[name = string("op_8006_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8006_cast_fp16 = reduce_mean(axes = var_8006_axes_0, keep_dims = var_8006_keep_dims_0, x = var_8004_cast_fp16)[name = string("op_8006_cast_fp16")];
            fp16 var_8007_to_fp16 = const()[name = string("op_8007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = add(x = var_8006_cast_fp16, y = var_8007_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = rsqrt(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8010_cast_fp16 = mul(x = x_263_cast_fp16, y = var_8009_cast_fp16)[name = string("op_8010_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648492096)))];
            tensor<fp16, [1, 16, 1024]> input_215_cast_fp16 = mul(x = var_8010_cast_fp16, y = model_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648494208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651640000))))[name = string("model_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8015_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8015_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651646208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654792000))))[name = string("model_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_219_cast_fp16 = mul(x = var_8015_cast_fp16, y = linear_152_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654798208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657944000))))[name = string("model_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_265_cast_fp16 = add(x = x_263_cast_fp16, y = linear_153_cast_fp16)[name = string("x_265_cast_fp16")];
            fp16 var_8029_promoted_to_fp16 = const()[name = string("op_8029_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8031_cast_fp16 = pow(x = x_265_cast_fp16, y = var_8029_promoted_to_fp16)[name = string("op_8031_cast_fp16")];
            tensor<int32, [1]> var_8033_axes_0 = const()[name = string("op_8033_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8033_keep_dims_0 = const()[name = string("op_8033_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8033_cast_fp16 = reduce_mean(axes = var_8033_axes_0, keep_dims = var_8033_keep_dims_0, x = var_8031_cast_fp16)[name = string("op_8033_cast_fp16")];
            fp16 var_8034_to_fp16 = const()[name = string("op_8034_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8035_cast_fp16 = add(x = var_8033_cast_fp16, y = var_8034_to_fp16)[name = string("op_8035_cast_fp16")];
            fp32 var_8036_epsilon_0 = const()[name = string("op_8036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8036_cast_fp16 = rsqrt(epsilon = var_8036_epsilon_0, x = var_8035_cast_fp16)[name = string("op_8036_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8037_cast_fp16 = mul(x = x_265_cast_fp16, y = var_8036_cast_fp16)[name = string("op_8037_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657946112)))];
            tensor<fp16, [1, 16, 1024]> input_221_cast_fp16 = mul(x = var_8037_cast_fp16, y = model_layers_22_input_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657948224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660045440))))[name = string("model_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_8046 = const()[name = string("op_8046"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = reshape(shape = var_8046, x = linear_154_cast_fp16)[name = string("op_8047_cast_fp16")];
            tensor<int32, [4]> x_267_perm_0 = const()[name = string("x_267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8055_promoted_to_fp16 = const()[name = string("op_8055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = transpose(perm = x_267_perm_0, x = var_8047_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8057_cast_fp16 = pow(x = x_267_cast_fp16, y = var_8055_promoted_to_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [1]> var_8059_axes_0 = const()[name = string("op_8059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8059_keep_dims_0 = const()[name = string("op_8059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8059_cast_fp16 = reduce_mean(axes = var_8059_axes_0, keep_dims = var_8059_keep_dims_0, x = var_8057_cast_fp16)[name = string("op_8059_cast_fp16")];
            fp16 var_8060_to_fp16 = const()[name = string("op_8060_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8061_cast_fp16 = add(x = var_8059_cast_fp16, y = var_8060_to_fp16)[name = string("op_8061_cast_fp16")];
            fp32 var_8062_epsilon_0 = const()[name = string("op_8062_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8062_cast_fp16 = rsqrt(epsilon = var_8062_epsilon_0, x = var_8061_cast_fp16)[name = string("op_8062_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8063_cast_fp16 = mul(x = x_267_cast_fp16, y = var_8062_cast_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049600)))];
            tensor<fp16, [1, 16, 16, 128]> x_271_cast_fp16 = mul(x = var_8063_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661098560))))[name = string("model_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_8072 = const()[name = string("op_8072"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8073_cast_fp16 = reshape(shape = var_8072, x = linear_155_cast_fp16)[name = string("op_8073_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8081_promoted_to_fp16 = const()[name = string("op_8081_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_8073_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8083_cast_fp16 = pow(x = x_269_cast_fp16, y = var_8081_promoted_to_fp16)[name = string("op_8083_cast_fp16")];
            tensor<int32, [1]> var_8085_axes_0 = const()[name = string("op_8085_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8085_keep_dims_0 = const()[name = string("op_8085_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8085_cast_fp16 = reduce_mean(axes = var_8085_axes_0, keep_dims = var_8085_keep_dims_0, x = var_8083_cast_fp16)[name = string("op_8085_cast_fp16")];
            fp16 var_8086_to_fp16 = const()[name = string("op_8086_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8087_cast_fp16 = add(x = var_8085_cast_fp16, y = var_8086_to_fp16)[name = string("op_8087_cast_fp16")];
            fp32 var_8088_epsilon_0 = const()[name = string("op_8088_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8088_cast_fp16 = rsqrt(epsilon = var_8088_epsilon_0, x = var_8087_cast_fp16)[name = string("op_8088_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8089_cast_fp16 = mul(x = x_269_cast_fp16, y = var_8088_cast_fp16)[name = string("op_8089_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100672)))];
            tensor<fp16, [1, 8, 16, 128]> x_273_cast_fp16 = mul(x = var_8089_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662149632))))[name = string("model_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_8098 = const()[name = string("op_8098"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8099_cast_fp16 = reshape(shape = var_8098, x = linear_156_cast_fp16)[name = string("op_8099_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_89_begin_0 = const()[name = string("first_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_89_end_0 = const()[name = string("first_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_89_end_mask_0 = const()[name = string("first_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_89_cast_fp16 = slice_by_index(begin = first_89_begin_0, end = first_89_end_0, end_mask = first_89_end_mask_0, x = x_271_cast_fp16)[name = string("first_89_cast_fp16")];
            tensor<int32, [4]> second_89_begin_0 = const()[name = string("second_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_89_end_0 = const()[name = string("second_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_89_end_mask_0 = const()[name = string("second_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_89_cast_fp16 = slice_by_index(begin = second_89_begin_0, end = second_89_end_0, end_mask = second_89_end_mask_0, x = x_271_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8113_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8114_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8114_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8116_cast_fp16 = sub(x = var_8113_cast_fp16, y = var_8114_cast_fp16)[name = string("op_8116_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8117_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8117_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8118_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8118_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8120_cast_fp16 = add(x = var_8117_cast_fp16, y = var_8118_cast_fp16)[name = string("op_8120_cast_fp16")];
            int32 var_8122 = const()[name = string("op_8122"), val = int32(-1)];
            bool q_45_interleave_0 = const()[name = string("q_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_45_cast_fp16 = concat(axis = var_8122, interleave = q_45_interleave_0, values = (var_8116_cast_fp16, var_8120_cast_fp16))[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> first_91_begin_0 = const()[name = string("first_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_91_end_0 = const()[name = string("first_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_91_end_mask_0 = const()[name = string("first_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_91_cast_fp16 = slice_by_index(begin = first_91_begin_0, end = first_91_end_0, end_mask = first_91_end_mask_0, x = x_273_cast_fp16)[name = string("first_91_cast_fp16")];
            tensor<int32, [4]> second_91_begin_0 = const()[name = string("second_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_91_end_0 = const()[name = string("second_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_91_end_mask_0 = const()[name = string("second_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_91_cast_fp16 = slice_by_index(begin = second_91_begin_0, end = second_91_end_0, end_mask = second_91_end_mask_0, x = x_273_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8134_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8134_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8135_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8135_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8137_cast_fp16 = sub(x = var_8134_cast_fp16, y = var_8135_cast_fp16)[name = string("op_8137_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8138_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8139_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8141_cast_fp16 = add(x = var_8138_cast_fp16, y = var_8139_cast_fp16)[name = string("op_8141_cast_fp16")];
            int32 var_8143 = const()[name = string("op_8143"), val = int32(-1)];
            bool k_45_interleave_0 = const()[name = string("k_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_45_cast_fp16 = concat(axis = var_8143, interleave = k_45_interleave_0, values = (var_8137_cast_fp16, var_8141_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8154_begin_0 = const()[name = string("op_8154_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8154_end_0 = const()[name = string("op_8154_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8154_end_mask_0 = const()[name = string("op_8154_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8154_stride_0 = const()[name = string("op_8154_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8154_cast_fp16 = slice_by_index(begin = var_8154_begin_0, end = var_8154_end_0, end_mask = var_8154_end_mask_0, stride = var_8154_stride_0, x = q_45_cast_fp16)[name = string("op_8154_cast_fp16")];
            tensor<int32, [4]> var_8164_begin_0 = const()[name = string("op_8164_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8164_end_0 = const()[name = string("op_8164_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8164_end_mask_0 = const()[name = string("op_8164_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8164_stride_0 = const()[name = string("op_8164_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8164_cast_fp16 = slice_by_index(begin = var_8164_begin_0, end = var_8164_end_0, end_mask = var_8164_end_mask_0, stride = var_8164_stride_0, x = q_45_cast_fp16)[name = string("op_8164_cast_fp16")];
            int32 var_8166 = const()[name = string("op_8166"), val = int32(2)];
            bool stacked_45_interleave_0 = const()[name = string("stacked_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_45_cast_fp16 = concat(axis = var_8166, interleave = stacked_45_interleave_0, values = (var_8154_cast_fp16, var_8164_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8172_begin_0 = const()[name = string("op_8172_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8172_end_0 = const()[name = string("op_8172_end_0"), val = tensor<int32, [4]>([23, 8, 4112, 128])];
            tensor<bool, [4]> var_8172_end_mask_0 = const()[name = string("op_8172_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8172_cast_fp16 = slice_by_index(begin = var_8172_begin_0, end = var_8172_end_0, end_mask = var_8172_end_mask_0, x = read_state_0)[name = string("op_8172_cast_fp16")];
            tensor<int32, [4]> var_8182_begin_0 = const()[name = string("op_8182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8182_end_0 = const()[name = string("op_8182_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8182_end_mask_0 = const()[name = string("op_8182_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8182_cast_fp16 = slice_by_index(begin = var_8182_begin_0, end = var_8182_end_0, end_mask = var_8182_end_mask_0, x = var_8172_cast_fp16)[name = string("op_8182_cast_fp16")];
            bool var_8191_transpose_x_1 = const()[name = string("op_8191_transpose_x_1"), val = bool(false)];
            bool var_8191_transpose_y_1 = const()[name = string("op_8191_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8191_cast_fp16 = matmul(transpose_x = var_8191_transpose_x_1, transpose_y = var_8191_transpose_y_1, x = var_8182_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<int32, [4]> var_8194_perm_0 = const()[name = string("op_8194_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8195_to_fp16 = const()[name = string("op_8195_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8194_cast_fp16 = transpose(perm = var_8194_perm_0, x = var_8191_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 4096]> var_8196_cast_fp16 = mul(x = var_8194_cast_fp16, y = var_8195_to_fp16)[name = string("op_8196_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_45_cast_fp16 = add(x = var_8196_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8202_transpose_x_1 = const()[name = string("op_8202_transpose_x_1"), val = bool(false)];
            bool var_8202_transpose_y_1 = const()[name = string("op_8202_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8202_cast_fp16 = matmul(transpose_x = var_8202_transpose_x_1, transpose_y = var_8202_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8202_cast_fp16")];
            fp16 var_8203_to_fp16 = const()[name = string("op_8203_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8204_cast_fp16 = mul(x = var_8202_cast_fp16, y = var_8203_to_fp16)[name = string("op_8204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8204_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8208 = const()[name = string("op_8208"), val = int32(-1)];
            bool var_8209_interleave_0 = const()[name = string("op_8209_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8209_cast_fp16 = concat(axis = var_8208, interleave = var_8209_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8209_cast_fp16")];
            int32 var_8210 = const()[name = string("op_8210"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_45_cast_fp16 = softmax(axis = var_8210, x = var_8209_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8217_begin_0 = const()[name = string("op_8217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8217_end_0 = const()[name = string("op_8217_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8217_end_mask_0 = const()[name = string("op_8217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8217_cast_fp16 = slice_by_index(begin = var_8217_begin_0, end = var_8217_end_0, end_mask = var_8217_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8217_cast_fp16")];
            tensor<int32, [4]> var_8222_begin_0 = const()[name = string("op_8222_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8222_end_0 = const()[name = string("op_8222_end_0"), val = tensor<int32, [4]>([23, 8, 4112, 128])];
            tensor<bool, [4]> var_8222_end_mask_0 = const()[name = string("op_8222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8222_cast_fp16 = slice_by_index(begin = var_8222_begin_0, end = var_8222_end_0, end_mask = var_8222_end_mask_0, x = read_state_1)[name = string("op_8222_cast_fp16")];
            tensor<int32, [4]> var_8232_begin_0 = const()[name = string("op_8232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8232_end_0 = const()[name = string("op_8232_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8232_end_mask_0 = const()[name = string("op_8232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8232_cast_fp16 = slice_by_index(begin = var_8232_begin_0, end = var_8232_end_0, end_mask = var_8232_end_mask_0, x = var_8222_cast_fp16)[name = string("op_8232_cast_fp16")];
            bool var_8238_transpose_x_0 = const()[name = string("op_8238_transpose_x_0"), val = bool(false)];
            bool var_8238_transpose_y_0 = const()[name = string("op_8238_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8238_cast_fp16 = matmul(transpose_x = var_8238_transpose_x_0, transpose_y = var_8238_transpose_y_0, x = var_8217_cast_fp16, y = var_8232_cast_fp16)[name = string("op_8238_cast_fp16")];
            tensor<int32, [4]> var_8243_begin_0 = const()[name = string("op_8243_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8243_end_0 = const()[name = string("op_8243_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8243_end_mask_0 = const()[name = string("op_8243_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8243_cast_fp16 = slice_by_index(begin = var_8243_begin_0, end = var_8243_end_0, end_mask = var_8243_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8243_cast_fp16")];
            bool var_8244_transpose_x_0 = const()[name = string("op_8244_transpose_x_0"), val = bool(false)];
            bool var_8244_transpose_y_0 = const()[name = string("op_8244_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8099_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8244_cast_fp16 = matmul(transpose_x = var_8244_transpose_x_0, transpose_y = var_8244_transpose_y_0, x = var_8243_cast_fp16, y = v_45_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8238_cast_fp16, y = var_8244_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8261_begin_0 = const()[name = string("op_8261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8261_end_0 = const()[name = string("op_8261_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8261_end_mask_0 = const()[name = string("op_8261_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8261_cast_fp16 = slice_by_index(begin = var_8261_begin_0, end = var_8261_end_0, end_mask = var_8261_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8261_cast_fp16")];
            tensor<int32, [4]> var_8276_begin_0 = const()[name = string("op_8276_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8276_end_0 = const()[name = string("op_8276_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8276_end_mask_0 = const()[name = string("op_8276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8276_cast_fp16 = slice_by_index(begin = var_8276_begin_0, end = var_8276_end_0, end_mask = var_8276_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8276_cast_fp16")];
            int32 var_8278 = const()[name = string("op_8278"), val = int32(1)];
            bool mixed_91_interleave_0 = const()[name = string("mixed_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_91_cast_fp16 = concat(axis = var_8278, interleave = mixed_91_interleave_0, values = (var_8261_cast_fp16, var_8276_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8282_perm_0 = const()[name = string("op_8282_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8286 = const()[name = string("op_8286"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8282_cast_fp16 = transpose(perm = var_8282_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8286, x = var_8282_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662151744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664248960))))[name = string("model_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_275_cast_fp16 = add(x = x_265_cast_fp16, y = linear_157_cast_fp16)[name = string("x_275_cast_fp16")];
            fp16 var_8297_promoted_to_fp16 = const()[name = string("op_8297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8299_cast_fp16 = pow(x = x_275_cast_fp16, y = var_8297_promoted_to_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [1]> var_8301_axes_0 = const()[name = string("op_8301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8301_keep_dims_0 = const()[name = string("op_8301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = reduce_mean(axes = var_8301_axes_0, keep_dims = var_8301_keep_dims_0, x = var_8299_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp16 var_8302_to_fp16 = const()[name = string("op_8302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8303_cast_fp16 = add(x = var_8301_cast_fp16, y = var_8302_to_fp16)[name = string("op_8303_cast_fp16")];
            fp32 var_8304_epsilon_0 = const()[name = string("op_8304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8304_cast_fp16 = rsqrt(epsilon = var_8304_epsilon_0, x = var_8303_cast_fp16)[name = string("op_8304_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8305_cast_fp16 = mul(x = x_275_cast_fp16, y = var_8304_cast_fp16)[name = string("op_8305_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664251072)))];
            tensor<fp16, [1, 16, 1024]> input_225_cast_fp16 = mul(x = var_8305_cast_fp16, y = model_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664253184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667398976))))[name = string("model_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8310_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8310_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667405184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670550976))))[name = string("model_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_229_cast_fp16 = mul(x = var_8310_cast_fp16, y = linear_159_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670557184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673702976))))[name = string("model_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_160_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_277_cast_fp16 = add(x = x_275_cast_fp16, y = linear_160_cast_fp16)[name = string("x_277_cast_fp16")];
            fp16 var_8324_promoted_to_fp16 = const()[name = string("op_8324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8326_cast_fp16 = pow(x = x_277_cast_fp16, y = var_8324_promoted_to_fp16)[name = string("op_8326_cast_fp16")];
            tensor<int32, [1]> var_8328_axes_0 = const()[name = string("op_8328_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8328_keep_dims_0 = const()[name = string("op_8328_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8328_cast_fp16 = reduce_mean(axes = var_8328_axes_0, keep_dims = var_8328_keep_dims_0, x = var_8326_cast_fp16)[name = string("op_8328_cast_fp16")];
            fp16 var_8329_to_fp16 = const()[name = string("op_8329_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = add(x = var_8328_cast_fp16, y = var_8329_to_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8332_cast_fp16 = mul(x = x_277_cast_fp16, y = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673705088)))];
            tensor<fp16, [1, 16, 1024]> input_231_cast_fp16 = mul(x = var_8332_cast_fp16, y = model_layers_23_input_layernorm_weight_to_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673707200))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675804416))))[name = string("model_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_8341 = const()[name = string("op_8341"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8342_cast_fp16 = reshape(shape = var_8341, x = linear_161_cast_fp16)[name = string("op_8342_cast_fp16")];
            tensor<int32, [4]> x_279_perm_0 = const()[name = string("x_279_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8350_promoted_to_fp16 = const()[name = string("op_8350_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_279_cast_fp16 = transpose(perm = x_279_perm_0, x = var_8342_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_8352_cast_fp16 = pow(x = x_279_cast_fp16, y = var_8350_promoted_to_fp16)[name = string("op_8352_cast_fp16")];
            tensor<int32, [1]> var_8354_axes_0 = const()[name = string("op_8354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8354_keep_dims_0 = const()[name = string("op_8354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8354_cast_fp16 = reduce_mean(axes = var_8354_axes_0, keep_dims = var_8354_keep_dims_0, x = var_8352_cast_fp16)[name = string("op_8354_cast_fp16")];
            fp16 var_8355_to_fp16 = const()[name = string("op_8355_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8356_cast_fp16 = add(x = var_8354_cast_fp16, y = var_8355_to_fp16)[name = string("op_8356_cast_fp16")];
            fp32 var_8357_epsilon_0 = const()[name = string("op_8357_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8357_cast_fp16 = rsqrt(epsilon = var_8357_epsilon_0, x = var_8356_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8358_cast_fp16 = mul(x = x_279_cast_fp16, y = var_8357_cast_fp16)[name = string("op_8358_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808576)))];
            tensor<fp16, [1, 16, 16, 128]> x_283_cast_fp16 = mul(x = var_8358_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676857536))))[name = string("model_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_8367 = const()[name = string("op_8367"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8368_cast_fp16 = reshape(shape = var_8367, x = linear_162_cast_fp16)[name = string("op_8368_cast_fp16")];
            tensor<int32, [4]> x_281_perm_0 = const()[name = string("x_281_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8376_promoted_to_fp16 = const()[name = string("op_8376_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_281_cast_fp16 = transpose(perm = x_281_perm_0, x = var_8368_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_8378_cast_fp16 = pow(x = x_281_cast_fp16, y = var_8376_promoted_to_fp16)[name = string("op_8378_cast_fp16")];
            tensor<int32, [1]> var_8380_axes_0 = const()[name = string("op_8380_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8380_keep_dims_0 = const()[name = string("op_8380_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8380_cast_fp16 = reduce_mean(axes = var_8380_axes_0, keep_dims = var_8380_keep_dims_0, x = var_8378_cast_fp16)[name = string("op_8380_cast_fp16")];
            fp16 var_8381_to_fp16 = const()[name = string("op_8381_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8382_cast_fp16 = add(x = var_8380_cast_fp16, y = var_8381_to_fp16)[name = string("op_8382_cast_fp16")];
            fp32 var_8383_epsilon_0 = const()[name = string("op_8383_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8383_cast_fp16 = rsqrt(epsilon = var_8383_epsilon_0, x = var_8382_cast_fp16)[name = string("op_8383_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8384_cast_fp16 = mul(x = x_281_cast_fp16, y = var_8383_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859648)))];
            tensor<fp16, [1, 8, 16, 128]> x_285_cast_fp16 = mul(x = var_8384_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677908608))))[name = string("model_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8393 = const()[name = string("op_8393"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8394_cast_fp16 = reshape(shape = var_8393, x = linear_163_cast_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_93_begin_0 = const()[name = string("first_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_93_end_0 = const()[name = string("first_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_93_end_mask_0 = const()[name = string("first_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_93_cast_fp16 = slice_by_index(begin = first_93_begin_0, end = first_93_end_0, end_mask = first_93_end_mask_0, x = x_283_cast_fp16)[name = string("first_93_cast_fp16")];
            tensor<int32, [4]> second_93_begin_0 = const()[name = string("second_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_93_end_0 = const()[name = string("second_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_93_end_mask_0 = const()[name = string("second_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_93_cast_fp16 = slice_by_index(begin = second_93_begin_0, end = second_93_end_0, end_mask = second_93_end_mask_0, x = x_283_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8408_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8408_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8409_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8409_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8411_cast_fp16 = sub(x = var_8408_cast_fp16, y = var_8409_cast_fp16)[name = string("op_8411_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8412_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8412_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8413_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8415_cast_fp16 = add(x = var_8412_cast_fp16, y = var_8413_cast_fp16)[name = string("op_8415_cast_fp16")];
            int32 var_8417 = const()[name = string("op_8417"), val = int32(-1)];
            bool q_47_interleave_0 = const()[name = string("q_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_47_cast_fp16 = concat(axis = var_8417, interleave = q_47_interleave_0, values = (var_8411_cast_fp16, var_8415_cast_fp16))[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> first_95_begin_0 = const()[name = string("first_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_95_end_0 = const()[name = string("first_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_95_end_mask_0 = const()[name = string("first_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_95_cast_fp16 = slice_by_index(begin = first_95_begin_0, end = first_95_end_0, end_mask = first_95_end_mask_0, x = x_285_cast_fp16)[name = string("first_95_cast_fp16")];
            tensor<int32, [4]> second_95_begin_0 = const()[name = string("second_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_95_end_0 = const()[name = string("second_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_95_end_mask_0 = const()[name = string("second_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_95_cast_fp16 = slice_by_index(begin = second_95_begin_0, end = second_95_end_0, end_mask = second_95_end_mask_0, x = x_285_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8429_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8429_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8430_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8432_cast_fp16 = sub(x = var_8429_cast_fp16, y = var_8430_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8433_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8434_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8434_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8436_cast_fp16 = add(x = var_8433_cast_fp16, y = var_8434_cast_fp16)[name = string("op_8436_cast_fp16")];
            int32 var_8438 = const()[name = string("op_8438"), val = int32(-1)];
            bool k_47_interleave_0 = const()[name = string("k_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_47_cast_fp16 = concat(axis = var_8438, interleave = k_47_interleave_0, values = (var_8432_cast_fp16, var_8436_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_8449_begin_0 = const()[name = string("op_8449_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8449_end_0 = const()[name = string("op_8449_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8449_end_mask_0 = const()[name = string("op_8449_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8449_stride_0 = const()[name = string("op_8449_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8449_cast_fp16 = slice_by_index(begin = var_8449_begin_0, end = var_8449_end_0, end_mask = var_8449_end_mask_0, stride = var_8449_stride_0, x = q_47_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<int32, [4]> var_8459_begin_0 = const()[name = string("op_8459_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8459_end_0 = const()[name = string("op_8459_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8459_end_mask_0 = const()[name = string("op_8459_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8459_stride_0 = const()[name = string("op_8459_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8459_cast_fp16 = slice_by_index(begin = var_8459_begin_0, end = var_8459_end_0, end_mask = var_8459_end_mask_0, stride = var_8459_stride_0, x = q_47_cast_fp16)[name = string("op_8459_cast_fp16")];
            int32 var_8461 = const()[name = string("op_8461"), val = int32(2)];
            bool stacked_47_interleave_0 = const()[name = string("stacked_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_47_cast_fp16 = concat(axis = var_8461, interleave = stacked_47_interleave_0, values = (var_8449_cast_fp16, var_8459_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_8467_begin_0 = const()[name = string("op_8467_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8467_end_0 = const()[name = string("op_8467_end_0"), val = tensor<int32, [4]>([24, 8, 4112, 128])];
            tensor<bool, [4]> var_8467_end_mask_0 = const()[name = string("op_8467_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8467_cast_fp16 = slice_by_index(begin = var_8467_begin_0, end = var_8467_end_0, end_mask = var_8467_end_mask_0, x = read_state_0)[name = string("op_8467_cast_fp16")];
            tensor<int32, [4]> var_8477_begin_0 = const()[name = string("op_8477_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8477_end_0 = const()[name = string("op_8477_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8477_end_mask_0 = const()[name = string("op_8477_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8477_cast_fp16 = slice_by_index(begin = var_8477_begin_0, end = var_8477_end_0, end_mask = var_8477_end_mask_0, x = var_8467_cast_fp16)[name = string("op_8477_cast_fp16")];
            bool var_8486_transpose_x_1 = const()[name = string("op_8486_transpose_x_1"), val = bool(false)];
            bool var_8486_transpose_y_1 = const()[name = string("op_8486_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8486_cast_fp16 = matmul(transpose_x = var_8486_transpose_x_1, transpose_y = var_8486_transpose_y_1, x = var_8477_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8489_perm_0 = const()[name = string("op_8489_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8490_to_fp16 = const()[name = string("op_8490_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8489_cast_fp16 = transpose(perm = var_8489_perm_0, x = var_8486_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 4096]> var_8491_cast_fp16 = mul(x = var_8489_cast_fp16, y = var_8490_to_fp16)[name = string("op_8491_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_47_cast_fp16 = add(x = var_8491_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_8497_transpose_x_1 = const()[name = string("op_8497_transpose_x_1"), val = bool(false)];
            bool var_8497_transpose_y_1 = const()[name = string("op_8497_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8497_cast_fp16 = matmul(transpose_x = var_8497_transpose_x_1, transpose_y = var_8497_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_8497_cast_fp16")];
            fp16 var_8498_to_fp16 = const()[name = string("op_8498_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8499_cast_fp16 = mul(x = var_8497_cast_fp16, y = var_8498_to_fp16)[name = string("op_8499_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_8499_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_8503 = const()[name = string("op_8503"), val = int32(-1)];
            bool var_8504_interleave_0 = const()[name = string("op_8504_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8504_cast_fp16 = concat(axis = var_8503, interleave = var_8504_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_8504_cast_fp16")];
            int32 var_8505 = const()[name = string("op_8505"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_47_cast_fp16 = softmax(axis = var_8505, x = var_8504_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_8512_begin_0 = const()[name = string("op_8512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8512_end_0 = const()[name = string("op_8512_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8512_end_mask_0 = const()[name = string("op_8512_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8512_cast_fp16 = slice_by_index(begin = var_8512_begin_0, end = var_8512_end_0, end_mask = var_8512_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8512_cast_fp16")];
            tensor<int32, [4]> var_8517_begin_0 = const()[name = string("op_8517_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8517_end_0 = const()[name = string("op_8517_end_0"), val = tensor<int32, [4]>([24, 8, 4112, 128])];
            tensor<bool, [4]> var_8517_end_mask_0 = const()[name = string("op_8517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8517_cast_fp16 = slice_by_index(begin = var_8517_begin_0, end = var_8517_end_0, end_mask = var_8517_end_mask_0, x = read_state_1)[name = string("op_8517_cast_fp16")];
            tensor<int32, [4]> var_8527_begin_0 = const()[name = string("op_8527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8527_end_0 = const()[name = string("op_8527_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8527_end_mask_0 = const()[name = string("op_8527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8527_cast_fp16 = slice_by_index(begin = var_8527_begin_0, end = var_8527_end_0, end_mask = var_8527_end_mask_0, x = var_8517_cast_fp16)[name = string("op_8527_cast_fp16")];
            bool var_8533_transpose_x_0 = const()[name = string("op_8533_transpose_x_0"), val = bool(false)];
            bool var_8533_transpose_y_0 = const()[name = string("op_8533_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8533_cast_fp16 = matmul(transpose_x = var_8533_transpose_x_0, transpose_y = var_8533_transpose_y_0, x = var_8512_cast_fp16, y = var_8527_cast_fp16)[name = string("op_8533_cast_fp16")];
            tensor<int32, [4]> var_8538_begin_0 = const()[name = string("op_8538_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8538_end_0 = const()[name = string("op_8538_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8538_end_mask_0 = const()[name = string("op_8538_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8538_cast_fp16 = slice_by_index(begin = var_8538_begin_0, end = var_8538_end_0, end_mask = var_8538_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8538_cast_fp16")];
            bool var_8539_transpose_x_0 = const()[name = string("op_8539_transpose_x_0"), val = bool(false)];
            bool var_8539_transpose_y_0 = const()[name = string("op_8539_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_8394_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_8539_cast_fp16 = matmul(transpose_x = var_8539_transpose_x_0, transpose_y = var_8539_transpose_y_0, x = var_8538_cast_fp16, y = v_47_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_8533_cast_fp16, y = var_8539_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_8556_begin_0 = const()[name = string("op_8556_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8556_end_0 = const()[name = string("op_8556_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8556_end_mask_0 = const()[name = string("op_8556_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8556_cast_fp16 = slice_by_index(begin = var_8556_begin_0, end = var_8556_end_0, end_mask = var_8556_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8556_cast_fp16")];
            tensor<int32, [4]> var_8571_begin_0 = const()[name = string("op_8571_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8571_end_0 = const()[name = string("op_8571_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8571_end_mask_0 = const()[name = string("op_8571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8571_cast_fp16 = slice_by_index(begin = var_8571_begin_0, end = var_8571_end_0, end_mask = var_8571_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8571_cast_fp16")];
            int32 var_8573 = const()[name = string("op_8573"), val = int32(1)];
            bool mixed_95_interleave_0 = const()[name = string("mixed_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_95_cast_fp16 = concat(axis = var_8573, interleave = mixed_95_interleave_0, values = (var_8556_cast_fp16, var_8571_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_8577_perm_0 = const()[name = string("op_8577_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8581 = const()[name = string("op_8581"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8577_cast_fp16 = transpose(perm = var_8577_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_8581, x = var_8577_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677910720))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680007936))))[name = string("model_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_164_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_287_cast_fp16 = add(x = x_277_cast_fp16, y = linear_164_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_8592_promoted_to_fp16 = const()[name = string("op_8592_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8594_cast_fp16 = pow(x = x_287_cast_fp16, y = var_8592_promoted_to_fp16)[name = string("op_8594_cast_fp16")];
            tensor<int32, [1]> var_8596_axes_0 = const()[name = string("op_8596_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8596_keep_dims_0 = const()[name = string("op_8596_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8596_cast_fp16 = reduce_mean(axes = var_8596_axes_0, keep_dims = var_8596_keep_dims_0, x = var_8594_cast_fp16)[name = string("op_8596_cast_fp16")];
            fp16 var_8597_to_fp16 = const()[name = string("op_8597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8598_cast_fp16 = add(x = var_8596_cast_fp16, y = var_8597_to_fp16)[name = string("op_8598_cast_fp16")];
            fp32 var_8599_epsilon_0 = const()[name = string("op_8599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8599_cast_fp16 = rsqrt(epsilon = var_8599_epsilon_0, x = var_8598_cast_fp16)[name = string("op_8599_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8600_cast_fp16 = mul(x = x_287_cast_fp16, y = var_8599_cast_fp16)[name = string("op_8600_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680010048)))];
            tensor<fp16, [1, 16, 1024]> input_235_cast_fp16 = mul(x = var_8600_cast_fp16, y = model_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680012160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683157952))))[name = string("model_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8605_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_8605_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683164160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686309952))))[name = string("model_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_239_cast_fp16 = mul(x = var_8605_cast_fp16, y = linear_166_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686316160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689461952))))[name = string("model_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_167_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_289_cast_fp16 = add(x = x_287_cast_fp16, y = linear_167_cast_fp16)[name = string("x_289_cast_fp16")];
            fp16 var_8619_promoted_to_fp16 = const()[name = string("op_8619_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8621_cast_fp16 = pow(x = x_289_cast_fp16, y = var_8619_promoted_to_fp16)[name = string("op_8621_cast_fp16")];
            tensor<int32, [1]> var_8623_axes_0 = const()[name = string("op_8623_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8623_keep_dims_0 = const()[name = string("op_8623_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = reduce_mean(axes = var_8623_axes_0, keep_dims = var_8623_keep_dims_0, x = var_8621_cast_fp16)[name = string("op_8623_cast_fp16")];
            fp16 var_8624_to_fp16 = const()[name = string("op_8624_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_8624_to_fp16)[name = string("op_8625_cast_fp16")];
            fp32 var_8626_epsilon_0 = const()[name = string("op_8626_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8626_cast_fp16 = rsqrt(epsilon = var_8626_epsilon_0, x = var_8625_cast_fp16)[name = string("op_8626_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8627_cast_fp16 = mul(x = x_289_cast_fp16, y = var_8626_cast_fp16)[name = string("op_8627_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689464064)))];
            tensor<fp16, [1, 16, 1024]> input_241_cast_fp16 = mul(x = var_8627_cast_fp16, y = model_layers_24_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_24_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689466176))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691563392))))[name = string("model_layers_24_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_24_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [4]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8637_cast_fp16 = reshape(shape = var_8636, x = linear_168_cast_fp16)[name = string("op_8637_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8645_promoted_to_fp16 = const()[name = string("op_8645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_8637_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_8647_cast_fp16 = pow(x = x_291_cast_fp16, y = var_8645_promoted_to_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> var_8649_axes_0 = const()[name = string("op_8649_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8649_keep_dims_0 = const()[name = string("op_8649_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8649_cast_fp16 = reduce_mean(axes = var_8649_axes_0, keep_dims = var_8649_keep_dims_0, x = var_8647_cast_fp16)[name = string("op_8649_cast_fp16")];
            fp16 var_8650_to_fp16 = const()[name = string("op_8650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8651_cast_fp16 = add(x = var_8649_cast_fp16, y = var_8650_to_fp16)[name = string("op_8651_cast_fp16")];
            fp32 var_8652_epsilon_0 = const()[name = string("op_8652_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8652_cast_fp16 = rsqrt(epsilon = var_8652_epsilon_0, x = var_8651_cast_fp16)[name = string("op_8652_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8653_cast_fp16 = mul(x = x_291_cast_fp16, y = var_8652_cast_fp16)[name = string("op_8653_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567552)))];
            tensor<fp16, [1, 16, 16, 128]> x_295_cast_fp16 = mul(x = var_8653_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_295_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692616512))))[name = string("model_layers_24_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [4]> var_8662 = const()[name = string("op_8662"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8663_cast_fp16 = reshape(shape = var_8662, x = linear_169_cast_fp16)[name = string("op_8663_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8671_promoted_to_fp16 = const()[name = string("op_8671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_8663_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_8673_cast_fp16 = pow(x = x_293_cast_fp16, y = var_8671_promoted_to_fp16)[name = string("op_8673_cast_fp16")];
            tensor<int32, [1]> var_8675_axes_0 = const()[name = string("op_8675_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8675_keep_dims_0 = const()[name = string("op_8675_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8675_cast_fp16 = reduce_mean(axes = var_8675_axes_0, keep_dims = var_8675_keep_dims_0, x = var_8673_cast_fp16)[name = string("op_8675_cast_fp16")];
            fp16 var_8676_to_fp16 = const()[name = string("op_8676_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8677_cast_fp16 = add(x = var_8675_cast_fp16, y = var_8676_to_fp16)[name = string("op_8677_cast_fp16")];
            fp32 var_8678_epsilon_0 = const()[name = string("op_8678_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8678_cast_fp16 = rsqrt(epsilon = var_8678_epsilon_0, x = var_8677_cast_fp16)[name = string("op_8678_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8679_cast_fp16 = mul(x = x_293_cast_fp16, y = var_8678_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618624)))];
            tensor<fp16, [1, 8, 16, 128]> x_297_cast_fp16 = mul(x = var_8679_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_297_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693667584))))[name = string("model_layers_24_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> var_8688 = const()[name = string("op_8688"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8689_cast_fp16 = reshape(shape = var_8688, x = linear_170_cast_fp16)[name = string("op_8689_cast_fp16")];
            tensor<int32, [4]> v_49_perm_0 = const()[name = string("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_97_begin_0 = const()[name = string("first_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_97_end_0 = const()[name = string("first_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_97_end_mask_0 = const()[name = string("first_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_97_cast_fp16 = slice_by_index(begin = first_97_begin_0, end = first_97_end_0, end_mask = first_97_end_mask_0, x = x_295_cast_fp16)[name = string("first_97_cast_fp16")];
            tensor<int32, [4]> second_97_begin_0 = const()[name = string("second_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_97_end_0 = const()[name = string("second_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_97_end_mask_0 = const()[name = string("second_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_97_cast_fp16 = slice_by_index(begin = second_97_begin_0, end = second_97_end_0, end_mask = second_97_end_mask_0, x = x_295_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8703_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8704_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8704_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8706_cast_fp16 = sub(x = var_8703_cast_fp16, y = var_8704_cast_fp16)[name = string("op_8706_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8707_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8707_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8708_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8708_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8710_cast_fp16 = add(x = var_8707_cast_fp16, y = var_8708_cast_fp16)[name = string("op_8710_cast_fp16")];
            int32 var_8712 = const()[name = string("op_8712"), val = int32(-1)];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_49_cast_fp16 = concat(axis = var_8712, interleave = q_49_interleave_0, values = (var_8706_cast_fp16, var_8710_cast_fp16))[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> first_99_begin_0 = const()[name = string("first_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_99_end_0 = const()[name = string("first_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_99_end_mask_0 = const()[name = string("first_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_99_cast_fp16 = slice_by_index(begin = first_99_begin_0, end = first_99_end_0, end_mask = first_99_end_mask_0, x = x_297_cast_fp16)[name = string("first_99_cast_fp16")];
            tensor<int32, [4]> second_99_begin_0 = const()[name = string("second_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_99_end_0 = const()[name = string("second_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_99_end_mask_0 = const()[name = string("second_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_99_cast_fp16 = slice_by_index(begin = second_99_begin_0, end = second_99_end_0, end_mask = second_99_end_mask_0, x = x_297_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8724_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8724_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8725_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8727_cast_fp16 = sub(x = var_8724_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8727_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8728_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8729_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8729_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8731_cast_fp16 = add(x = var_8728_cast_fp16, y = var_8729_cast_fp16)[name = string("op_8731_cast_fp16")];
            int32 var_8733 = const()[name = string("op_8733"), val = int32(-1)];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_49_cast_fp16 = concat(axis = var_8733, interleave = k_49_interleave_0, values = (var_8727_cast_fp16, var_8731_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_8744_begin_0 = const()[name = string("op_8744_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8744_end_0 = const()[name = string("op_8744_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8744_end_mask_0 = const()[name = string("op_8744_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8744_stride_0 = const()[name = string("op_8744_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, stride = var_8744_stride_0, x = q_49_cast_fp16)[name = string("op_8744_cast_fp16")];
            tensor<int32, [4]> var_8754_begin_0 = const()[name = string("op_8754_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8754_end_0 = const()[name = string("op_8754_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8754_end_mask_0 = const()[name = string("op_8754_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8754_stride_0 = const()[name = string("op_8754_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8754_cast_fp16 = slice_by_index(begin = var_8754_begin_0, end = var_8754_end_0, end_mask = var_8754_end_mask_0, stride = var_8754_stride_0, x = q_49_cast_fp16)[name = string("op_8754_cast_fp16")];
            int32 var_8756 = const()[name = string("op_8756"), val = int32(2)];
            bool stacked_49_interleave_0 = const()[name = string("stacked_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_49_cast_fp16 = concat(axis = var_8756, interleave = stacked_49_interleave_0, values = (var_8744_cast_fp16, var_8754_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_8762_begin_0 = const()[name = string("op_8762_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8762_end_0 = const()[name = string("op_8762_end_0"), val = tensor<int32, [4]>([25, 8, 4112, 128])];
            tensor<bool, [4]> var_8762_end_mask_0 = const()[name = string("op_8762_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8762_cast_fp16 = slice_by_index(begin = var_8762_begin_0, end = var_8762_end_0, end_mask = var_8762_end_mask_0, x = read_state_0)[name = string("op_8762_cast_fp16")];
            tensor<int32, [4]> var_8772_begin_0 = const()[name = string("op_8772_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8772_end_0 = const()[name = string("op_8772_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8772_end_mask_0 = const()[name = string("op_8772_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8772_cast_fp16 = slice_by_index(begin = var_8772_begin_0, end = var_8772_end_0, end_mask = var_8772_end_mask_0, x = var_8762_cast_fp16)[name = string("op_8772_cast_fp16")];
            bool var_8781_transpose_x_1 = const()[name = string("op_8781_transpose_x_1"), val = bool(false)];
            bool var_8781_transpose_y_1 = const()[name = string("op_8781_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_8781_cast_fp16 = matmul(transpose_x = var_8781_transpose_x_1, transpose_y = var_8781_transpose_y_1, x = var_8772_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_8781_cast_fp16")];
            tensor<int32, [4]> var_8784_perm_0 = const()[name = string("op_8784_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8785_to_fp16 = const()[name = string("op_8785_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_8784_cast_fp16 = transpose(perm = var_8784_perm_0, x = var_8781_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 4096]> var_8786_cast_fp16 = mul(x = var_8784_cast_fp16, y = var_8785_to_fp16)[name = string("op_8786_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_49_cast_fp16 = add(x = var_8786_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_8792_transpose_x_1 = const()[name = string("op_8792_transpose_x_1"), val = bool(false)];
            bool var_8792_transpose_y_1 = const()[name = string("op_8792_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8792_cast_fp16 = matmul(transpose_x = var_8792_transpose_x_1, transpose_y = var_8792_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_8792_cast_fp16")];
            fp16 var_8793_to_fp16 = const()[name = string("op_8793_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8794_cast_fp16 = mul(x = var_8792_cast_fp16, y = var_8793_to_fp16)[name = string("op_8794_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_8794_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_8798 = const()[name = string("op_8798"), val = int32(-1)];
            bool var_8799_interleave_0 = const()[name = string("op_8799_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_8799_cast_fp16 = concat(axis = var_8798, interleave = var_8799_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_8799_cast_fp16")];
            int32 var_8800 = const()[name = string("op_8800"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_49_cast_fp16 = softmax(axis = var_8800, x = var_8799_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_8807_begin_0 = const()[name = string("op_8807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8807_end_0 = const()[name = string("op_8807_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_8807_end_mask_0 = const()[name = string("op_8807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_8807_cast_fp16 = slice_by_index(begin = var_8807_begin_0, end = var_8807_end_0, end_mask = var_8807_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8807_cast_fp16")];
            tensor<int32, [4]> var_8812_begin_0 = const()[name = string("op_8812_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8812_end_0 = const()[name = string("op_8812_end_0"), val = tensor<int32, [4]>([25, 8, 4112, 128])];
            tensor<bool, [4]> var_8812_end_mask_0 = const()[name = string("op_8812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_8812_cast_fp16 = slice_by_index(begin = var_8812_begin_0, end = var_8812_end_0, end_mask = var_8812_end_mask_0, x = read_state_1)[name = string("op_8812_cast_fp16")];
            tensor<int32, [4]> var_8822_begin_0 = const()[name = string("op_8822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8822_end_0 = const()[name = string("op_8822_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_8822_end_mask_0 = const()[name = string("op_8822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_8822_cast_fp16 = slice_by_index(begin = var_8822_begin_0, end = var_8822_end_0, end_mask = var_8822_end_mask_0, x = var_8812_cast_fp16)[name = string("op_8822_cast_fp16")];
            bool var_8828_transpose_x_0 = const()[name = string("op_8828_transpose_x_0"), val = bool(false)];
            bool var_8828_transpose_y_0 = const()[name = string("op_8828_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_0, transpose_y = var_8828_transpose_y_0, x = var_8807_cast_fp16, y = var_8822_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8833_begin_0 = const()[name = string("op_8833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_8833_end_0 = const()[name = string("op_8833_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_8833_end_mask_0 = const()[name = string("op_8833_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8833_cast_fp16 = slice_by_index(begin = var_8833_begin_0, end = var_8833_end_0, end_mask = var_8833_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8833_cast_fp16")];
            bool var_8834_transpose_x_0 = const()[name = string("op_8834_transpose_x_0"), val = bool(false)];
            bool var_8834_transpose_y_0 = const()[name = string("op_8834_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_8689_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_8834_cast_fp16 = matmul(transpose_x = var_8834_transpose_x_0, transpose_y = var_8834_transpose_y_0, x = var_8833_cast_fp16, y = v_49_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_8828_cast_fp16, y = var_8834_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_8851_begin_0 = const()[name = string("op_8851_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8851_end_0 = const()[name = string("op_8851_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8851_end_mask_0 = const()[name = string("op_8851_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8851_cast_fp16 = slice_by_index(begin = var_8851_begin_0, end = var_8851_end_0, end_mask = var_8851_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<int32, [4]> var_8866_begin_0 = const()[name = string("op_8866_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8866_end_0 = const()[name = string("op_8866_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8866_end_mask_0 = const()[name = string("op_8866_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8866_cast_fp16 = slice_by_index(begin = var_8866_begin_0, end = var_8866_end_0, end_mask = var_8866_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8866_cast_fp16")];
            int32 var_8868 = const()[name = string("op_8868"), val = int32(1)];
            bool mixed_99_interleave_0 = const()[name = string("mixed_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_99_cast_fp16 = concat(axis = var_8868, interleave = mixed_99_interleave_0, values = (var_8851_cast_fp16, var_8866_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_8872_perm_0 = const()[name = string("op_8872_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8876 = const()[name = string("op_8876"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8872_cast_fp16 = transpose(perm = var_8872_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_8876, x = var_8872_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_24_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693669696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695766912))))[name = string("model_layers_24_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_171_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = linear_171_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_8887_promoted_to_fp16 = const()[name = string("op_8887_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8889_cast_fp16 = pow(x = x_299_cast_fp16, y = var_8887_promoted_to_fp16)[name = string("op_8889_cast_fp16")];
            tensor<int32, [1]> var_8891_axes_0 = const()[name = string("op_8891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8891_keep_dims_0 = const()[name = string("op_8891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8891_cast_fp16 = reduce_mean(axes = var_8891_axes_0, keep_dims = var_8891_keep_dims_0, x = var_8889_cast_fp16)[name = string("op_8891_cast_fp16")];
            fp16 var_8892_to_fp16 = const()[name = string("op_8892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8893_cast_fp16 = add(x = var_8891_cast_fp16, y = var_8892_to_fp16)[name = string("op_8893_cast_fp16")];
            fp32 var_8894_epsilon_0 = const()[name = string("op_8894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8894_cast_fp16 = rsqrt(epsilon = var_8894_epsilon_0, x = var_8893_cast_fp16)[name = string("op_8894_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8895_cast_fp16 = mul(x = x_299_cast_fp16, y = var_8894_cast_fp16)[name = string("op_8895_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695769024)))];
            tensor<fp16, [1, 16, 1024]> input_245_cast_fp16 = mul(x = var_8895_cast_fp16, y = model_layers_24_post_attention_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695771136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698916928))))[name = string("model_layers_24_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8900_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_8900_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698923136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702068928))))[name = string("model_layers_24_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_249_cast_fp16 = mul(x = var_8900_cast_fp16, y = linear_173_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_24_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702075136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705220928))))[name = string("model_layers_24_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_174_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_mlp_down_proj_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_301_cast_fp16 = add(x = x_299_cast_fp16, y = linear_174_cast_fp16)[name = string("x_301_cast_fp16")];
            fp16 var_8914_promoted_to_fp16 = const()[name = string("op_8914_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8916_cast_fp16 = pow(x = x_301_cast_fp16, y = var_8914_promoted_to_fp16)[name = string("op_8916_cast_fp16")];
            tensor<int32, [1]> var_8918_axes_0 = const()[name = string("op_8918_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8918_keep_dims_0 = const()[name = string("op_8918_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8918_cast_fp16 = reduce_mean(axes = var_8918_axes_0, keep_dims = var_8918_keep_dims_0, x = var_8916_cast_fp16)[name = string("op_8918_cast_fp16")];
            fp16 var_8919_to_fp16 = const()[name = string("op_8919_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8920_cast_fp16 = add(x = var_8918_cast_fp16, y = var_8919_to_fp16)[name = string("op_8920_cast_fp16")];
            fp32 var_8921_epsilon_0 = const()[name = string("op_8921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8921_cast_fp16 = rsqrt(epsilon = var_8921_epsilon_0, x = var_8920_cast_fp16)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8922_cast_fp16 = mul(x = x_301_cast_fp16, y = var_8921_cast_fp16)[name = string("op_8922_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705223040)))];
            tensor<fp16, [1, 16, 1024]> input_251_cast_fp16 = mul(x = var_8922_cast_fp16, y = model_layers_25_input_layernorm_weight_to_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_25_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705225152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707322368))))[name = string("model_layers_25_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_25_self_attn_q_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> var_8931 = const()[name = string("op_8931"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8932_cast_fp16 = reshape(shape = var_8931, x = linear_175_cast_fp16)[name = string("op_8932_cast_fp16")];
            tensor<int32, [4]> x_303_perm_0 = const()[name = string("x_303_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8940_promoted_to_fp16 = const()[name = string("op_8940_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_303_cast_fp16 = transpose(perm = x_303_perm_0, x = var_8932_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_8942_cast_fp16 = pow(x = x_303_cast_fp16, y = var_8940_promoted_to_fp16)[name = string("op_8942_cast_fp16")];
            tensor<int32, [1]> var_8944_axes_0 = const()[name = string("op_8944_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8944_keep_dims_0 = const()[name = string("op_8944_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8944_cast_fp16 = reduce_mean(axes = var_8944_axes_0, keep_dims = var_8944_keep_dims_0, x = var_8942_cast_fp16)[name = string("op_8944_cast_fp16")];
            fp16 var_8945_to_fp16 = const()[name = string("op_8945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8946_cast_fp16 = add(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            fp32 var_8947_epsilon_0 = const()[name = string("op_8947_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8947_cast_fp16 = rsqrt(epsilon = var_8947_epsilon_0, x = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8948_cast_fp16 = mul(x = x_303_cast_fp16, y = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326528)))];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = mul(x = var_8948_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708375488))))[name = string("model_layers_25_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_176_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_k_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> var_8957 = const()[name = string("op_8957"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8958_cast_fp16 = reshape(shape = var_8957, x = linear_176_cast_fp16)[name = string("op_8958_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_8958_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_8968_cast_fp16 = pow(x = x_305_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8968_cast_fp16")];
            tensor<int32, [1]> var_8970_axes_0 = const()[name = string("op_8970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8970_keep_dims_0 = const()[name = string("op_8970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8970_cast_fp16 = reduce_mean(axes = var_8970_axes_0, keep_dims = var_8970_keep_dims_0, x = var_8968_cast_fp16)[name = string("op_8970_cast_fp16")];
            fp16 var_8971_to_fp16 = const()[name = string("op_8971_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8972_cast_fp16 = add(x = var_8970_cast_fp16, y = var_8971_to_fp16)[name = string("op_8972_cast_fp16")];
            fp32 var_8973_epsilon_0 = const()[name = string("op_8973_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8973_cast_fp16 = rsqrt(epsilon = var_8973_epsilon_0, x = var_8972_cast_fp16)[name = string("op_8973_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8974_cast_fp16 = mul(x = x_305_cast_fp16, y = var_8973_cast_fp16)[name = string("op_8974_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377600)))];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = mul(x = var_8974_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_309_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709426560))))[name = string("model_layers_25_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_v_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> var_8983 = const()[name = string("op_8983"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8984_cast_fp16 = reshape(shape = var_8983, x = linear_177_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_101_begin_0 = const()[name = string("first_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_101_end_0 = const()[name = string("first_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_101_end_mask_0 = const()[name = string("first_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_101_cast_fp16 = slice_by_index(begin = first_101_begin_0, end = first_101_end_0, end_mask = first_101_end_mask_0, x = x_307_cast_fp16)[name = string("first_101_cast_fp16")];
            tensor<int32, [4]> second_101_begin_0 = const()[name = string("second_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_101_end_0 = const()[name = string("second_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_101_end_mask_0 = const()[name = string("second_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_101_cast_fp16 = slice_by_index(begin = second_101_begin_0, end = second_101_end_0, end_mask = second_101_end_mask_0, x = x_307_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8998_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8998_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8999_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8999_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9001_cast_fp16 = sub(x = var_8998_cast_fp16, y = var_8999_cast_fp16)[name = string("op_9001_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9002_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9002_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9003_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9005_cast_fp16 = add(x = var_9002_cast_fp16, y = var_9003_cast_fp16)[name = string("op_9005_cast_fp16")];
            int32 var_9007 = const()[name = string("op_9007"), val = int32(-1)];
            bool q_51_interleave_0 = const()[name = string("q_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_51_cast_fp16 = concat(axis = var_9007, interleave = q_51_interleave_0, values = (var_9001_cast_fp16, var_9005_cast_fp16))[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> first_103_begin_0 = const()[name = string("first_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_103_end_0 = const()[name = string("first_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_103_end_mask_0 = const()[name = string("first_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_103_cast_fp16 = slice_by_index(begin = first_103_begin_0, end = first_103_end_0, end_mask = first_103_end_mask_0, x = x_309_cast_fp16)[name = string("first_103_cast_fp16")];
            tensor<int32, [4]> second_103_begin_0 = const()[name = string("second_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_103_end_0 = const()[name = string("second_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_103_end_mask_0 = const()[name = string("second_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_103_cast_fp16 = slice_by_index(begin = second_103_begin_0, end = second_103_end_0, end_mask = second_103_end_mask_0, x = x_309_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9019_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9019_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9020_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9020_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9022_cast_fp16 = sub(x = var_9019_cast_fp16, y = var_9020_cast_fp16)[name = string("op_9022_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9023_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9024_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9024_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9026_cast_fp16 = add(x = var_9023_cast_fp16, y = var_9024_cast_fp16)[name = string("op_9026_cast_fp16")];
            int32 var_9028 = const()[name = string("op_9028"), val = int32(-1)];
            bool k_51_interleave_0 = const()[name = string("k_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_51_cast_fp16 = concat(axis = var_9028, interleave = k_51_interleave_0, values = (var_9022_cast_fp16, var_9026_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9039_begin_0 = const()[name = string("op_9039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9039_end_0 = const()[name = string("op_9039_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9039_end_mask_0 = const()[name = string("op_9039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9039_stride_0 = const()[name = string("op_9039_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9039_cast_fp16 = slice_by_index(begin = var_9039_begin_0, end = var_9039_end_0, end_mask = var_9039_end_mask_0, stride = var_9039_stride_0, x = q_51_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [4]> var_9049_begin_0 = const()[name = string("op_9049_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9049_end_0 = const()[name = string("op_9049_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9049_end_mask_0 = const()[name = string("op_9049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9049_stride_0 = const()[name = string("op_9049_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = slice_by_index(begin = var_9049_begin_0, end = var_9049_end_0, end_mask = var_9049_end_mask_0, stride = var_9049_stride_0, x = q_51_cast_fp16)[name = string("op_9049_cast_fp16")];
            int32 var_9051 = const()[name = string("op_9051"), val = int32(2)];
            bool stacked_51_interleave_0 = const()[name = string("stacked_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_51_cast_fp16 = concat(axis = var_9051, interleave = stacked_51_interleave_0, values = (var_9039_cast_fp16, var_9049_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9057_begin_0 = const()[name = string("op_9057_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9057_end_0 = const()[name = string("op_9057_end_0"), val = tensor<int32, [4]>([26, 8, 4112, 128])];
            tensor<bool, [4]> var_9057_end_mask_0 = const()[name = string("op_9057_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9057_cast_fp16 = slice_by_index(begin = var_9057_begin_0, end = var_9057_end_0, end_mask = var_9057_end_mask_0, x = read_state_0)[name = string("op_9057_cast_fp16")];
            tensor<int32, [4]> var_9067_begin_0 = const()[name = string("op_9067_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9067_end_0 = const()[name = string("op_9067_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9067_end_mask_0 = const()[name = string("op_9067_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9067_cast_fp16 = slice_by_index(begin = var_9067_begin_0, end = var_9067_end_0, end_mask = var_9067_end_mask_0, x = var_9057_cast_fp16)[name = string("op_9067_cast_fp16")];
            bool var_9076_transpose_x_1 = const()[name = string("op_9076_transpose_x_1"), val = bool(false)];
            bool var_9076_transpose_y_1 = const()[name = string("op_9076_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9076_cast_fp16 = matmul(transpose_x = var_9076_transpose_x_1, transpose_y = var_9076_transpose_y_1, x = var_9067_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<int32, [4]> var_9079_perm_0 = const()[name = string("op_9079_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9080_to_fp16 = const()[name = string("op_9080_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9079_cast_fp16 = transpose(perm = var_9079_perm_0, x = var_9076_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 4096]> var_9081_cast_fp16 = mul(x = var_9079_cast_fp16, y = var_9080_to_fp16)[name = string("op_9081_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_51_cast_fp16 = add(x = var_9081_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9087_transpose_x_1 = const()[name = string("op_9087_transpose_x_1"), val = bool(false)];
            bool var_9087_transpose_y_1 = const()[name = string("op_9087_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9087_cast_fp16 = matmul(transpose_x = var_9087_transpose_x_1, transpose_y = var_9087_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9089_cast_fp16 = mul(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("op_9089_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9089_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9093 = const()[name = string("op_9093"), val = int32(-1)];
            bool var_9094_interleave_0 = const()[name = string("op_9094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9094_cast_fp16 = concat(axis = var_9093, interleave = var_9094_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9094_cast_fp16")];
            int32 var_9095 = const()[name = string("op_9095"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_51_cast_fp16 = softmax(axis = var_9095, x = var_9094_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9102_begin_0 = const()[name = string("op_9102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9102_end_0 = const()[name = string("op_9102_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9102_end_mask_0 = const()[name = string("op_9102_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9102_cast_fp16 = slice_by_index(begin = var_9102_begin_0, end = var_9102_end_0, end_mask = var_9102_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9102_cast_fp16")];
            tensor<int32, [4]> var_9107_begin_0 = const()[name = string("op_9107_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9107_end_0 = const()[name = string("op_9107_end_0"), val = tensor<int32, [4]>([26, 8, 4112, 128])];
            tensor<bool, [4]> var_9107_end_mask_0 = const()[name = string("op_9107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9107_cast_fp16 = slice_by_index(begin = var_9107_begin_0, end = var_9107_end_0, end_mask = var_9107_end_mask_0, x = read_state_1)[name = string("op_9107_cast_fp16")];
            tensor<int32, [4]> var_9117_begin_0 = const()[name = string("op_9117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9117_end_0 = const()[name = string("op_9117_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9117_end_mask_0 = const()[name = string("op_9117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9117_cast_fp16 = slice_by_index(begin = var_9117_begin_0, end = var_9117_end_0, end_mask = var_9117_end_mask_0, x = var_9107_cast_fp16)[name = string("op_9117_cast_fp16")];
            bool var_9123_transpose_x_0 = const()[name = string("op_9123_transpose_x_0"), val = bool(false)];
            bool var_9123_transpose_y_0 = const()[name = string("op_9123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9123_cast_fp16 = matmul(transpose_x = var_9123_transpose_x_0, transpose_y = var_9123_transpose_y_0, x = var_9102_cast_fp16, y = var_9117_cast_fp16)[name = string("op_9123_cast_fp16")];
            tensor<int32, [4]> var_9128_begin_0 = const()[name = string("op_9128_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9128_end_0 = const()[name = string("op_9128_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9128_end_mask_0 = const()[name = string("op_9128_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9128_cast_fp16 = slice_by_index(begin = var_9128_begin_0, end = var_9128_end_0, end_mask = var_9128_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9128_cast_fp16")];
            bool var_9129_transpose_x_0 = const()[name = string("op_9129_transpose_x_0"), val = bool(false)];
            bool var_9129_transpose_y_0 = const()[name = string("op_9129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_8984_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9129_cast_fp16 = matmul(transpose_x = var_9129_transpose_x_0, transpose_y = var_9129_transpose_y_0, x = var_9128_cast_fp16, y = v_51_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9123_cast_fp16, y = var_9129_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9146_begin_0 = const()[name = string("op_9146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9146_end_0 = const()[name = string("op_9146_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9146_end_mask_0 = const()[name = string("op_9146_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9146_cast_fp16 = slice_by_index(begin = var_9146_begin_0, end = var_9146_end_0, end_mask = var_9146_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9146_cast_fp16")];
            tensor<int32, [4]> var_9161_begin_0 = const()[name = string("op_9161_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9161_end_0 = const()[name = string("op_9161_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9161_end_mask_0 = const()[name = string("op_9161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9161_cast_fp16 = slice_by_index(begin = var_9161_begin_0, end = var_9161_end_0, end_mask = var_9161_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9161_cast_fp16")];
            int32 var_9163 = const()[name = string("op_9163"), val = int32(1)];
            bool mixed_103_interleave_0 = const()[name = string("mixed_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_103_cast_fp16 = concat(axis = var_9163, interleave = mixed_103_interleave_0, values = (var_9146_cast_fp16, var_9161_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9167_perm_0 = const()[name = string("op_9167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9171 = const()[name = string("op_9171"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9167_cast_fp16 = transpose(perm = var_9167_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9171, x = var_9167_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_25_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709428672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711525888))))[name = string("model_layers_25_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_178_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_o_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_311_cast_fp16 = add(x = x_301_cast_fp16, y = linear_178_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_9182_promoted_to_fp16 = const()[name = string("op_9182_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9184_cast_fp16 = pow(x = x_311_cast_fp16, y = var_9182_promoted_to_fp16)[name = string("op_9184_cast_fp16")];
            tensor<int32, [1]> var_9186_axes_0 = const()[name = string("op_9186_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9186_keep_dims_0 = const()[name = string("op_9186_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9186_cast_fp16 = reduce_mean(axes = var_9186_axes_0, keep_dims = var_9186_keep_dims_0, x = var_9184_cast_fp16)[name = string("op_9186_cast_fp16")];
            fp16 var_9187_to_fp16 = const()[name = string("op_9187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9188_cast_fp16 = add(x = var_9186_cast_fp16, y = var_9187_to_fp16)[name = string("op_9188_cast_fp16")];
            fp32 var_9189_epsilon_0 = const()[name = string("op_9189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9189_cast_fp16 = rsqrt(epsilon = var_9189_epsilon_0, x = var_9188_cast_fp16)[name = string("op_9189_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9190_cast_fp16 = mul(x = x_311_cast_fp16, y = var_9189_cast_fp16)[name = string("op_9190_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711528000)))];
            tensor<fp16, [1, 16, 1024]> input_255_cast_fp16 = mul(x = var_9190_cast_fp16, y = model_layers_25_post_attention_layernorm_weight_to_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711530112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714675904))))[name = string("model_layers_25_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_gate_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9195_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714682112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717827904))))[name = string("model_layers_25_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_up_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_259_cast_fp16 = mul(x = var_9195_cast_fp16, y = linear_180_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_25_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717834112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720979904))))[name = string("model_layers_25_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_313_cast_fp16 = add(x = x_311_cast_fp16, y = linear_181_cast_fp16)[name = string("x_313_cast_fp16")];
            fp16 var_9209_promoted_to_fp16 = const()[name = string("op_9209_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9211_cast_fp16 = pow(x = x_313_cast_fp16, y = var_9209_promoted_to_fp16)[name = string("op_9211_cast_fp16")];
            tensor<int32, [1]> var_9213_axes_0 = const()[name = string("op_9213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9213_keep_dims_0 = const()[name = string("op_9213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9213_cast_fp16 = reduce_mean(axes = var_9213_axes_0, keep_dims = var_9213_keep_dims_0, x = var_9211_cast_fp16)[name = string("op_9213_cast_fp16")];
            fp16 var_9214_to_fp16 = const()[name = string("op_9214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9215_cast_fp16 = add(x = var_9213_cast_fp16, y = var_9214_to_fp16)[name = string("op_9215_cast_fp16")];
            fp32 var_9216_epsilon_0 = const()[name = string("op_9216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9216_cast_fp16 = rsqrt(epsilon = var_9216_epsilon_0, x = var_9215_cast_fp16)[name = string("op_9216_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9217_cast_fp16 = mul(x = x_313_cast_fp16, y = var_9216_cast_fp16)[name = string("op_9217_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720982016)))];
            tensor<fp16, [1, 16, 1024]> input_261_cast_fp16 = mul(x = var_9217_cast_fp16, y = model_layers_26_input_layernorm_weight_to_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_26_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720984128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723081344))))[name = string("model_layers_26_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_26_self_attn_q_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_9226 = const()[name = string("op_9226"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9227_cast_fp16 = reshape(shape = var_9226, x = linear_182_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<int32, [4]> x_315_perm_0 = const()[name = string("x_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9235_promoted_to_fp16 = const()[name = string("op_9235_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_315_cast_fp16 = transpose(perm = x_315_perm_0, x = var_9227_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9237_cast_fp16 = pow(x = x_315_cast_fp16, y = var_9235_promoted_to_fp16)[name = string("op_9237_cast_fp16")];
            tensor<int32, [1]> var_9239_axes_0 = const()[name = string("op_9239_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9239_keep_dims_0 = const()[name = string("op_9239_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9239_cast_fp16 = reduce_mean(axes = var_9239_axes_0, keep_dims = var_9239_keep_dims_0, x = var_9237_cast_fp16)[name = string("op_9239_cast_fp16")];
            fp16 var_9240_to_fp16 = const()[name = string("op_9240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9241_cast_fp16 = add(x = var_9239_cast_fp16, y = var_9240_to_fp16)[name = string("op_9241_cast_fp16")];
            fp32 var_9242_epsilon_0 = const()[name = string("op_9242_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9242_cast_fp16 = rsqrt(epsilon = var_9242_epsilon_0, x = var_9241_cast_fp16)[name = string("op_9242_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9243_cast_fp16 = mul(x = x_315_cast_fp16, y = var_9242_cast_fp16)[name = string("op_9243_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085504)))];
            tensor<fp16, [1, 16, 16, 128]> x_319_cast_fp16 = mul(x = var_9243_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_319_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724134464))))[name = string("model_layers_26_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_183_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_k_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_9252 = const()[name = string("op_9252"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9253_cast_fp16 = reshape(shape = var_9252, x = linear_183_cast_fp16)[name = string("op_9253_cast_fp16")];
            tensor<int32, [4]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9261_promoted_to_fp16 = const()[name = string("op_9261_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = var_9253_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_9263_cast_fp16 = pow(x = x_317_cast_fp16, y = var_9261_promoted_to_fp16)[name = string("op_9263_cast_fp16")];
            tensor<int32, [1]> var_9265_axes_0 = const()[name = string("op_9265_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9265_keep_dims_0 = const()[name = string("op_9265_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9265_cast_fp16 = reduce_mean(axes = var_9265_axes_0, keep_dims = var_9265_keep_dims_0, x = var_9263_cast_fp16)[name = string("op_9265_cast_fp16")];
            fp16 var_9266_to_fp16 = const()[name = string("op_9266_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9267_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9266_to_fp16)[name = string("op_9267_cast_fp16")];
            fp32 var_9268_epsilon_0 = const()[name = string("op_9268_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9268_cast_fp16 = rsqrt(epsilon = var_9268_epsilon_0, x = var_9267_cast_fp16)[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9269_cast_fp16 = mul(x = x_317_cast_fp16, y = var_9268_cast_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136576)))];
            tensor<fp16, [1, 8, 16, 128]> x_321_cast_fp16 = mul(x = var_9269_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_321_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725185536))))[name = string("model_layers_26_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_184_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_v_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_9278 = const()[name = string("op_9278"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9279_cast_fp16 = reshape(shape = var_9278, x = linear_184_cast_fp16)[name = string("op_9279_cast_fp16")];
            tensor<int32, [4]> v_53_perm_0 = const()[name = string("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_105_begin_0 = const()[name = string("first_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_105_end_0 = const()[name = string("first_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_105_end_mask_0 = const()[name = string("first_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_105_cast_fp16 = slice_by_index(begin = first_105_begin_0, end = first_105_end_0, end_mask = first_105_end_mask_0, x = x_319_cast_fp16)[name = string("first_105_cast_fp16")];
            tensor<int32, [4]> second_105_begin_0 = const()[name = string("second_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_105_end_0 = const()[name = string("second_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_105_end_mask_0 = const()[name = string("second_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_105_cast_fp16 = slice_by_index(begin = second_105_begin_0, end = second_105_end_0, end_mask = second_105_end_mask_0, x = x_319_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9293_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9294_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9296_cast_fp16 = sub(x = var_9293_cast_fp16, y = var_9294_cast_fp16)[name = string("op_9296_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9297_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9297_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9298_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9298_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9300_cast_fp16 = add(x = var_9297_cast_fp16, y = var_9298_cast_fp16)[name = string("op_9300_cast_fp16")];
            int32 var_9302 = const()[name = string("op_9302"), val = int32(-1)];
            bool q_53_interleave_0 = const()[name = string("q_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_53_cast_fp16 = concat(axis = var_9302, interleave = q_53_interleave_0, values = (var_9296_cast_fp16, var_9300_cast_fp16))[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> first_107_begin_0 = const()[name = string("first_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_107_end_0 = const()[name = string("first_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_107_end_mask_0 = const()[name = string("first_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_107_cast_fp16 = slice_by_index(begin = first_107_begin_0, end = first_107_end_0, end_mask = first_107_end_mask_0, x = x_321_cast_fp16)[name = string("first_107_cast_fp16")];
            tensor<int32, [4]> second_107_begin_0 = const()[name = string("second_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_107_end_0 = const()[name = string("second_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_107_end_mask_0 = const()[name = string("second_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_107_cast_fp16 = slice_by_index(begin = second_107_begin_0, end = second_107_end_0, end_mask = second_107_end_mask_0, x = x_321_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9314_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9315_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9317_cast_fp16 = sub(x = var_9314_cast_fp16, y = var_9315_cast_fp16)[name = string("op_9317_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9318_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9319_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9319_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9321_cast_fp16 = add(x = var_9318_cast_fp16, y = var_9319_cast_fp16)[name = string("op_9321_cast_fp16")];
            int32 var_9323 = const()[name = string("op_9323"), val = int32(-1)];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_53_cast_fp16 = concat(axis = var_9323, interleave = k_53_interleave_0, values = (var_9317_cast_fp16, var_9321_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_9334_begin_0 = const()[name = string("op_9334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9334_end_0 = const()[name = string("op_9334_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9334_end_mask_0 = const()[name = string("op_9334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9334_stride_0 = const()[name = string("op_9334_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9334_cast_fp16 = slice_by_index(begin = var_9334_begin_0, end = var_9334_end_0, end_mask = var_9334_end_mask_0, stride = var_9334_stride_0, x = q_53_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [4]> var_9344_begin_0 = const()[name = string("op_9344_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9344_end_0 = const()[name = string("op_9344_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9344_end_mask_0 = const()[name = string("op_9344_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9344_stride_0 = const()[name = string("op_9344_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9344_cast_fp16 = slice_by_index(begin = var_9344_begin_0, end = var_9344_end_0, end_mask = var_9344_end_mask_0, stride = var_9344_stride_0, x = q_53_cast_fp16)[name = string("op_9344_cast_fp16")];
            int32 var_9346 = const()[name = string("op_9346"), val = int32(2)];
            bool stacked_53_interleave_0 = const()[name = string("stacked_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_53_cast_fp16 = concat(axis = var_9346, interleave = stacked_53_interleave_0, values = (var_9334_cast_fp16, var_9344_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_9352_begin_0 = const()[name = string("op_9352_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9352_end_0 = const()[name = string("op_9352_end_0"), val = tensor<int32, [4]>([27, 8, 4112, 128])];
            tensor<bool, [4]> var_9352_end_mask_0 = const()[name = string("op_9352_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9352_cast_fp16 = slice_by_index(begin = var_9352_begin_0, end = var_9352_end_0, end_mask = var_9352_end_mask_0, x = read_state_0)[name = string("op_9352_cast_fp16")];
            tensor<int32, [4]> var_9362_begin_0 = const()[name = string("op_9362_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9362_end_0 = const()[name = string("op_9362_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9362_end_mask_0 = const()[name = string("op_9362_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9362_cast_fp16 = slice_by_index(begin = var_9362_begin_0, end = var_9362_end_0, end_mask = var_9362_end_mask_0, x = var_9352_cast_fp16)[name = string("op_9362_cast_fp16")];
            bool var_9371_transpose_x_1 = const()[name = string("op_9371_transpose_x_1"), val = bool(false)];
            bool var_9371_transpose_y_1 = const()[name = string("op_9371_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9371_cast_fp16 = matmul(transpose_x = var_9371_transpose_x_1, transpose_y = var_9371_transpose_y_1, x = var_9362_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<int32, [4]> var_9374_perm_0 = const()[name = string("op_9374_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9375_to_fp16 = const()[name = string("op_9375_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9374_cast_fp16 = transpose(perm = var_9374_perm_0, x = var_9371_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 4096]> var_9376_cast_fp16 = mul(x = var_9374_cast_fp16, y = var_9375_to_fp16)[name = string("op_9376_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_53_cast_fp16 = add(x = var_9376_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_9382_transpose_x_1 = const()[name = string("op_9382_transpose_x_1"), val = bool(false)];
            bool var_9382_transpose_y_1 = const()[name = string("op_9382_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9382_cast_fp16 = matmul(transpose_x = var_9382_transpose_x_1, transpose_y = var_9382_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_9382_cast_fp16")];
            fp16 var_9383_to_fp16 = const()[name = string("op_9383_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9384_cast_fp16 = mul(x = var_9382_cast_fp16, y = var_9383_to_fp16)[name = string("op_9384_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_9384_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_9388 = const()[name = string("op_9388"), val = int32(-1)];
            bool var_9389_interleave_0 = const()[name = string("op_9389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9389_cast_fp16 = concat(axis = var_9388, interleave = var_9389_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_9389_cast_fp16")];
            int32 var_9390 = const()[name = string("op_9390"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_53_cast_fp16 = softmax(axis = var_9390, x = var_9389_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_9397_begin_0 = const()[name = string("op_9397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9397_end_0 = const()[name = string("op_9397_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9397_end_mask_0 = const()[name = string("op_9397_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9397_cast_fp16 = slice_by_index(begin = var_9397_begin_0, end = var_9397_end_0, end_mask = var_9397_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<int32, [4]> var_9402_begin_0 = const()[name = string("op_9402_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9402_end_0 = const()[name = string("op_9402_end_0"), val = tensor<int32, [4]>([27, 8, 4112, 128])];
            tensor<bool, [4]> var_9402_end_mask_0 = const()[name = string("op_9402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9402_cast_fp16 = slice_by_index(begin = var_9402_begin_0, end = var_9402_end_0, end_mask = var_9402_end_mask_0, x = read_state_1)[name = string("op_9402_cast_fp16")];
            tensor<int32, [4]> var_9412_begin_0 = const()[name = string("op_9412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9412_end_0 = const()[name = string("op_9412_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9412_end_mask_0 = const()[name = string("op_9412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9412_cast_fp16 = slice_by_index(begin = var_9412_begin_0, end = var_9412_end_0, end_mask = var_9412_end_mask_0, x = var_9402_cast_fp16)[name = string("op_9412_cast_fp16")];
            bool var_9418_transpose_x_0 = const()[name = string("op_9418_transpose_x_0"), val = bool(false)];
            bool var_9418_transpose_y_0 = const()[name = string("op_9418_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9418_cast_fp16 = matmul(transpose_x = var_9418_transpose_x_0, transpose_y = var_9418_transpose_y_0, x = var_9397_cast_fp16, y = var_9412_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<int32, [4]> var_9423_begin_0 = const()[name = string("op_9423_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9423_end_0 = const()[name = string("op_9423_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9423_end_mask_0 = const()[name = string("op_9423_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9423_cast_fp16 = slice_by_index(begin = var_9423_begin_0, end = var_9423_end_0, end_mask = var_9423_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9423_cast_fp16")];
            bool var_9424_transpose_x_0 = const()[name = string("op_9424_transpose_x_0"), val = bool(false)];
            bool var_9424_transpose_y_0 = const()[name = string("op_9424_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_9279_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_9424_cast_fp16 = matmul(transpose_x = var_9424_transpose_x_0, transpose_y = var_9424_transpose_y_0, x = var_9423_cast_fp16, y = v_53_cast_fp16)[name = string("op_9424_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_9418_cast_fp16, y = var_9424_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_9441_begin_0 = const()[name = string("op_9441_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9441_end_0 = const()[name = string("op_9441_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9441_end_mask_0 = const()[name = string("op_9441_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9441_cast_fp16 = slice_by_index(begin = var_9441_begin_0, end = var_9441_end_0, end_mask = var_9441_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9441_cast_fp16")];
            tensor<int32, [4]> var_9456_begin_0 = const()[name = string("op_9456_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9456_end_0 = const()[name = string("op_9456_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9456_end_mask_0 = const()[name = string("op_9456_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9456_cast_fp16 = slice_by_index(begin = var_9456_begin_0, end = var_9456_end_0, end_mask = var_9456_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9456_cast_fp16")];
            int32 var_9458 = const()[name = string("op_9458"), val = int32(1)];
            bool mixed_107_interleave_0 = const()[name = string("mixed_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_107_cast_fp16 = concat(axis = var_9458, interleave = mixed_107_interleave_0, values = (var_9441_cast_fp16, var_9456_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_9462_perm_0 = const()[name = string("op_9462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9466 = const()[name = string("op_9466"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9462_cast_fp16 = transpose(perm = var_9462_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_9466, x = var_9462_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_26_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725187648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727284864))))[name = string("model_layers_26_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_o_proj_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = add(x = x_313_cast_fp16, y = linear_185_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_9477_promoted_to_fp16 = const()[name = string("op_9477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9479_cast_fp16 = pow(x = x_323_cast_fp16, y = var_9477_promoted_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<int32, [1]> var_9481_axes_0 = const()[name = string("op_9481_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9481_keep_dims_0 = const()[name = string("op_9481_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9481_cast_fp16 = reduce_mean(axes = var_9481_axes_0, keep_dims = var_9481_keep_dims_0, x = var_9479_cast_fp16)[name = string("op_9481_cast_fp16")];
            fp16 var_9482_to_fp16 = const()[name = string("op_9482_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9483_cast_fp16 = add(x = var_9481_cast_fp16, y = var_9482_to_fp16)[name = string("op_9483_cast_fp16")];
            fp32 var_9484_epsilon_0 = const()[name = string("op_9484_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9484_cast_fp16 = rsqrt(epsilon = var_9484_epsilon_0, x = var_9483_cast_fp16)[name = string("op_9484_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9485_cast_fp16 = mul(x = x_323_cast_fp16, y = var_9484_cast_fp16)[name = string("op_9485_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727286976)))];
            tensor<fp16, [1, 16, 1024]> input_265_cast_fp16 = mul(x = var_9485_cast_fp16, y = model_layers_26_post_attention_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727289088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730434880))))[name = string("model_layers_26_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_gate_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9490_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_9490_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730441088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733586880))))[name = string("model_layers_26_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_up_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_269_cast_fp16 = mul(x = var_9490_cast_fp16, y = linear_187_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_26_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733593088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736738880))))[name = string("model_layers_26_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_mlp_down_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_325_cast_fp16 = add(x = x_323_cast_fp16, y = linear_188_cast_fp16)[name = string("x_325_cast_fp16")];
            fp16 var_9504_promoted_to_fp16 = const()[name = string("op_9504_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9506_cast_fp16 = pow(x = x_325_cast_fp16, y = var_9504_promoted_to_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9508_axes_0 = const()[name = string("op_9508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9508_keep_dims_0 = const()[name = string("op_9508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9508_cast_fp16 = reduce_mean(axes = var_9508_axes_0, keep_dims = var_9508_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9508_cast_fp16")];
            fp16 var_9509_to_fp16 = const()[name = string("op_9509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9510_cast_fp16 = add(x = var_9508_cast_fp16, y = var_9509_to_fp16)[name = string("op_9510_cast_fp16")];
            fp32 var_9511_epsilon_0 = const()[name = string("op_9511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9511_cast_fp16 = rsqrt(epsilon = var_9511_epsilon_0, x = var_9510_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9512_cast_fp16 = mul(x = x_325_cast_fp16, y = var_9511_cast_fp16)[name = string("op_9512_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736740992)))];
            tensor<fp16, [1, 16, 1024]> input_271_cast_fp16 = mul(x = var_9512_cast_fp16, y = model_layers_27_input_layernorm_weight_to_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_27_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736743104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738840320))))[name = string("model_layers_27_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_27_self_attn_q_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<int32, [4]> var_9521 = const()[name = string("op_9521"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9522_cast_fp16 = reshape(shape = var_9521, x = linear_189_cast_fp16)[name = string("op_9522_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9530_promoted_to_fp16 = const()[name = string("op_9530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_9522_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_9532_cast_fp16 = pow(x = x_327_cast_fp16, y = var_9530_promoted_to_fp16)[name = string("op_9532_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9532_cast_fp16)[name = string("op_9534_cast_fp16")];
            fp16 var_9535_to_fp16 = const()[name = string("op_9535_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9536_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9535_to_fp16)[name = string("op_9536_cast_fp16")];
            fp32 var_9537_epsilon_0 = const()[name = string("op_9537_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9537_cast_fp16 = rsqrt(epsilon = var_9537_epsilon_0, x = var_9536_cast_fp16)[name = string("op_9537_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9538_cast_fp16 = mul(x = x_327_cast_fp16, y = var_9537_cast_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844480)))];
            tensor<fp16, [1, 16, 16, 128]> x_331_cast_fp16 = mul(x = var_9538_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_331_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739893440))))[name = string("model_layers_27_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_k_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9547 = const()[name = string("op_9547"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9548_cast_fp16 = reshape(shape = var_9547, x = linear_190_cast_fp16)[name = string("op_9548_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9556_promoted_to_fp16 = const()[name = string("op_9556_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_9548_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_9558_cast_fp16 = pow(x = x_329_cast_fp16, y = var_9556_promoted_to_fp16)[name = string("op_9558_cast_fp16")];
            tensor<int32, [1]> var_9560_axes_0 = const()[name = string("op_9560_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9560_keep_dims_0 = const()[name = string("op_9560_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9560_cast_fp16 = reduce_mean(axes = var_9560_axes_0, keep_dims = var_9560_keep_dims_0, x = var_9558_cast_fp16)[name = string("op_9560_cast_fp16")];
            fp16 var_9561_to_fp16 = const()[name = string("op_9561_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9562_cast_fp16 = add(x = var_9560_cast_fp16, y = var_9561_to_fp16)[name = string("op_9562_cast_fp16")];
            fp32 var_9563_epsilon_0 = const()[name = string("op_9563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9563_cast_fp16 = rsqrt(epsilon = var_9563_epsilon_0, x = var_9562_cast_fp16)[name = string("op_9563_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9564_cast_fp16 = mul(x = x_329_cast_fp16, y = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895552)))];
            tensor<fp16, [1, 8, 16, 128]> x_333_cast_fp16 = mul(x = var_9564_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_333_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740944512))))[name = string("model_layers_27_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_191_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_v_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9573 = const()[name = string("op_9573"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9574_cast_fp16 = reshape(shape = var_9573, x = linear_191_cast_fp16)[name = string("op_9574_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_109_begin_0 = const()[name = string("first_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_109_end_0 = const()[name = string("first_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_109_end_mask_0 = const()[name = string("first_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_109_cast_fp16 = slice_by_index(begin = first_109_begin_0, end = first_109_end_0, end_mask = first_109_end_mask_0, x = x_331_cast_fp16)[name = string("first_109_cast_fp16")];
            tensor<int32, [4]> second_109_begin_0 = const()[name = string("second_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_109_end_0 = const()[name = string("second_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_109_end_mask_0 = const()[name = string("second_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_109_cast_fp16 = slice_by_index(begin = second_109_begin_0, end = second_109_end_0, end_mask = second_109_end_mask_0, x = x_331_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9588_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9588_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9589_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9589_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9591_cast_fp16 = sub(x = var_9588_cast_fp16, y = var_9589_cast_fp16)[name = string("op_9591_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9592_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9593_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9593_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9595_cast_fp16 = add(x = var_9592_cast_fp16, y = var_9593_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_cast_fp16 = concat(axis = var_9597, interleave = q_interleave_0, values = (var_9591_cast_fp16, var_9595_cast_fp16))[name = string("q_cast_fp16")];
            tensor<int32, [4]> first_begin_0 = const()[name = string("first_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_end_0 = const()[name = string("first_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_end_mask_0 = const()[name = string("first_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_cast_fp16 = slice_by_index(begin = first_begin_0, end = first_end_0, end_mask = first_end_mask_0, x = x_333_cast_fp16)[name = string("first_cast_fp16")];
            tensor<int32, [4]> second_begin_0 = const()[name = string("second_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_end_0 = const()[name = string("second_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_end_mask_0 = const()[name = string("second_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_cast_fp16 = slice_by_index(begin = second_begin_0, end = second_end_0, end_mask = second_end_mask_0, x = x_333_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9609_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9609_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9610_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9610_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9612_cast_fp16 = sub(x = var_9609_cast_fp16, y = var_9610_cast_fp16)[name = string("op_9612_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9613_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9614_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9614_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9616_cast_fp16 = add(x = var_9613_cast_fp16, y = var_9614_cast_fp16)[name = string("op_9616_cast_fp16")];
            int32 var_9618 = const()[name = string("op_9618"), val = int32(-1)];
            bool k_interleave_0 = const()[name = string("k_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_cast_fp16 = concat(axis = var_9618, interleave = k_interleave_0, values = (var_9612_cast_fp16, var_9616_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_9629_begin_0 = const()[name = string("op_9629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9629_end_0 = const()[name = string("op_9629_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9629_end_mask_0 = const()[name = string("op_9629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9629_stride_0 = const()[name = string("op_9629_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9629_cast_fp16 = slice_by_index(begin = var_9629_begin_0, end = var_9629_end_0, end_mask = var_9629_end_mask_0, stride = var_9629_stride_0, x = q_cast_fp16)[name = string("op_9629_cast_fp16")];
            tensor<int32, [4]> var_9639_begin_0 = const()[name = string("op_9639_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9639_end_0 = const()[name = string("op_9639_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9639_end_mask_0 = const()[name = string("op_9639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9639_stride_0 = const()[name = string("op_9639_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9639_cast_fp16 = slice_by_index(begin = var_9639_begin_0, end = var_9639_end_0, end_mask = var_9639_end_mask_0, stride = var_9639_stride_0, x = q_cast_fp16)[name = string("op_9639_cast_fp16")];
            int32 var_9641 = const()[name = string("op_9641"), val = int32(2)];
            bool stacked_interleave_0 = const()[name = string("stacked_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_cast_fp16 = concat(axis = var_9641, interleave = stacked_interleave_0, values = (var_9629_cast_fp16, var_9639_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_9647_begin_0 = const()[name = string("op_9647_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9647_end_0 = const()[name = string("op_9647_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_9647_end_mask_0 = const()[name = string("op_9647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9647_cast_fp16 = slice_by_index(begin = var_9647_begin_0, end = var_9647_end_0, end_mask = var_9647_end_mask_0, x = read_state_0)[name = string("op_9647_cast_fp16")];
            tensor<int32, [4]> var_9657_begin_0 = const()[name = string("op_9657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9657_end_0 = const()[name = string("op_9657_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9657_end_mask_0 = const()[name = string("op_9657_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9657_cast_fp16 = slice_by_index(begin = var_9657_begin_0, end = var_9657_end_0, end_mask = var_9657_end_mask_0, x = var_9647_cast_fp16)[name = string("op_9657_cast_fp16")];
            bool var_9666_transpose_x_1 = const()[name = string("op_9666_transpose_x_1"), val = bool(false)];
            bool var_9666_transpose_y_1 = const()[name = string("op_9666_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 4096, 32]> var_9666_cast_fp16 = matmul(transpose_x = var_9666_transpose_x_1, transpose_y = var_9666_transpose_y_1, x = var_9657_cast_fp16, y = stacked_cast_fp16)[name = string("op_9666_cast_fp16")];
            tensor<int32, [4]> var_9669_perm_0 = const()[name = string("op_9669_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9670_to_fp16 = const()[name = string("op_9670_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 4096]> var_9669_cast_fp16 = transpose(perm = var_9669_perm_0, x = var_9666_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 4096]> var_9671_cast_fp16 = mul(x = var_9669_cast_fp16, y = var_9670_to_fp16)[name = string("op_9671_cast_fp16")];
            tensor<fp16, [1, 8, 32, 4096]> cache_weights_cast_fp16 = add(x = var_9671_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_9677_transpose_x_1 = const()[name = string("op_9677_transpose_x_1"), val = bool(false)];
            bool var_9677_transpose_y_1 = const()[name = string("op_9677_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9677_cast_fp16 = matmul(transpose_x = var_9677_transpose_x_1, transpose_y = var_9677_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_9677_cast_fp16")];
            fp16 var_9678_to_fp16 = const()[name = string("op_9678_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9679_cast_fp16 = mul(x = var_9677_cast_fp16, y = var_9678_to_fp16)[name = string("op_9679_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_9679_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_9683 = const()[name = string("op_9683"), val = int32(-1)];
            bool var_9684_interleave_0 = const()[name = string("op_9684_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 4112]> var_9684_cast_fp16 = concat(axis = var_9683, interleave = var_9684_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_9684_cast_fp16")];
            int32 var_9685 = const()[name = string("op_9685"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 4112]> shares_cast_fp16 = softmax(axis = var_9685, x = var_9684_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_9692_begin_0 = const()[name = string("op_9692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9692_end_0 = const()[name = string("op_9692_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4096])];
            tensor<bool, [4]> var_9692_end_mask_0 = const()[name = string("op_9692_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 4096]> var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, x = shares_cast_fp16)[name = string("op_9692_cast_fp16")];
            tensor<int32, [4]> var_9697_begin_0 = const()[name = string("op_9697_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9697_end_0 = const()[name = string("op_9697_end_0"), val = tensor<int32, [4]>([1, 8, 4112, 128])];
            tensor<bool, [4]> var_9697_end_mask_0 = const()[name = string("op_9697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 4112, 128]> var_9697_cast_fp16 = slice_by_index(begin = var_9697_begin_0, end = var_9697_end_0, end_mask = var_9697_end_mask_0, x = read_state_1)[name = string("op_9697_cast_fp16")];
            tensor<int32, [4]> var_9707_begin_0 = const()[name = string("op_9707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9707_end_0 = const()[name = string("op_9707_end_0"), val = tensor<int32, [4]>([1, 8, 4096, 128])];
            tensor<bool, [4]> var_9707_end_mask_0 = const()[name = string("op_9707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 4096, 128]> var_9707_cast_fp16 = slice_by_index(begin = var_9707_begin_0, end = var_9707_end_0, end_mask = var_9707_end_mask_0, x = var_9697_cast_fp16)[name = string("op_9707_cast_fp16")];
            bool var_9713_transpose_x_0 = const()[name = string("op_9713_transpose_x_0"), val = bool(false)];
            bool var_9713_transpose_y_0 = const()[name = string("op_9713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9713_cast_fp16 = matmul(transpose_x = var_9713_transpose_x_0, transpose_y = var_9713_transpose_y_0, x = var_9692_cast_fp16, y = var_9707_cast_fp16)[name = string("op_9713_cast_fp16")];
            tensor<int32, [4]> var_9718_begin_0 = const()[name = string("op_9718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 4096])];
            tensor<int32, [4]> var_9718_end_0 = const()[name = string("op_9718_end_0"), val = tensor<int32, [4]>([1, 8, 32, 4112])];
            tensor<bool, [4]> var_9718_end_mask_0 = const()[name = string("op_9718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9718_cast_fp16 = slice_by_index(begin = var_9718_begin_0, end = var_9718_end_0, end_mask = var_9718_end_mask_0, x = shares_cast_fp16)[name = string("op_9718_cast_fp16")];
            bool var_9719_transpose_x_0 = const()[name = string("op_9719_transpose_x_0"), val = bool(false)];
            bool var_9719_transpose_y_0 = const()[name = string("op_9719_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_9574_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_9719_cast_fp16 = matmul(transpose_x = var_9719_transpose_x_0, transpose_y = var_9719_transpose_y_0, x = var_9718_cast_fp16, y = v_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_9713_cast_fp16, y = var_9719_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_9736_begin_0 = const()[name = string("op_9736_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9736_end_0 = const()[name = string("op_9736_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9736_end_mask_0 = const()[name = string("op_9736_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9736_cast_fp16 = slice_by_index(begin = var_9736_begin_0, end = var_9736_end_0, end_mask = var_9736_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9736_cast_fp16")];
            tensor<int32, [4]> var_9751_begin_0 = const()[name = string("op_9751_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9751_end_0 = const()[name = string("op_9751_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9751_end_mask_0 = const()[name = string("op_9751_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9751_cast_fp16 = slice_by_index(begin = var_9751_begin_0, end = var_9751_end_0, end_mask = var_9751_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9751_cast_fp16")];
            int32 var_9753 = const()[name = string("op_9753"), val = int32(1)];
            bool mixed_interleave_0 = const()[name = string("mixed_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_cast_fp16 = concat(axis = var_9753, interleave = mixed_interleave_0, values = (var_9736_cast_fp16, var_9751_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_9757_perm_0 = const()[name = string("op_9757_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9761 = const()[name = string("op_9761"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9757_cast_fp16 = transpose(perm = var_9757_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_9761, x = var_9757_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_27_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740946624))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743043840))))[name = string("model_layers_27_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_192_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_o_proj_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = linear_192_cast_fp16)[name = string("x_335_cast_fp16")];
            fp16 var_9772_promoted_to_fp16 = const()[name = string("op_9772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9774_cast_fp16 = pow(x = x_335_cast_fp16, y = var_9772_promoted_to_fp16)[name = string("op_9774_cast_fp16")];
            tensor<int32, [1]> var_9776_axes_0 = const()[name = string("op_9776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9776_keep_dims_0 = const()[name = string("op_9776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9776_cast_fp16 = reduce_mean(axes = var_9776_axes_0, keep_dims = var_9776_keep_dims_0, x = var_9774_cast_fp16)[name = string("op_9776_cast_fp16")];
            fp16 var_9777_to_fp16 = const()[name = string("op_9777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9778_cast_fp16 = add(x = var_9776_cast_fp16, y = var_9777_to_fp16)[name = string("op_9778_cast_fp16")];
            fp32 var_9779_epsilon_0 = const()[name = string("op_9779_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9779_cast_fp16 = rsqrt(epsilon = var_9779_epsilon_0, x = var_9778_cast_fp16)[name = string("op_9779_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9780_cast_fp16 = mul(x = x_335_cast_fp16, y = var_9779_cast_fp16)[name = string("op_9780_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743045952)))];
            tensor<fp16, [1, 16, 1024]> input_275_cast_fp16 = mul(x = var_9780_cast_fp16, y = model_layers_27_post_attention_layernorm_weight_to_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743048064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746193856))))[name = string("model_layers_27_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_gate_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9785_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_9785_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746200064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749345856))))[name = string("model_layers_27_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_up_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_cast_fp16 = mul(x = var_9785_cast_fp16, y = linear_194_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_27_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749352064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752497856))))[name = string("model_layers_27_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_195_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_mlp_down_proj_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_337_cast_fp16 = add(x = x_335_cast_fp16, y = linear_195_cast_fp16)[name = string("x_337_cast_fp16")];
            int32 var_9795 = const()[name = string("op_9795"), val = int32(1)];
            int32 x_batch_dims_0 = const()[name = string("x_batch_dims_0"), val = int32(0)];
            bool x_validate_indices_0 = const()[name = string("x_validate_indices_0"), val = bool(false)];
            string answerRow_to_uint16_dtype_0 = const()[name = string("answerRow_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> answerRow_to_uint16 = cast(dtype = answerRow_to_uint16_dtype_0, x = answerRow)[name = string("cast_0")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16_cast_uint16 = gather(axis = var_9795, batch_dims = x_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_validate_indices_0, x = x_337_cast_fp16)[name = string("x_cast_fp16_cast_uint16")];
            fp16 var_9801_promoted_to_fp16 = const()[name = string("op_9801_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_9803_cast_fp16 = pow(x = x_cast_fp16_cast_uint16, y = var_9801_promoted_to_fp16)[name = string("op_9803_cast_fp16")];
            tensor<int32, [1]> var_9805_axes_0 = const()[name = string("op_9805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9805_keep_dims_0 = const()[name = string("op_9805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_9805_cast_fp16 = reduce_mean(axes = var_9805_axes_0, keep_dims = var_9805_keep_dims_0, x = var_9803_cast_fp16)[name = string("op_9805_cast_fp16")];
            fp16 var_9806_to_fp16 = const()[name = string("op_9806_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_9807_cast_fp16 = add(x = var_9805_cast_fp16, y = var_9806_to_fp16)[name = string("op_9807_cast_fp16")];
            fp32 var_9808_epsilon_0 = const()[name = string("op_9808_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_9808_cast_fp16 = rsqrt(epsilon = var_9808_epsilon_0, x = var_9807_cast_fp16)[name = string("op_9808_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_9809_cast_fp16 = mul(x = x_cast_fp16_cast_uint16, y = var_9808_cast_fp16)[name = string("op_9809_cast_fp16")];
            tensor<fp16, [1024]> model_norm_weight_to_fp16 = const()[name = string("model_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752499968)))];
            tensor<fp16, [1, 1, 1024]> last_cast_fp16 = mul(x = var_9809_cast_fp16, y = model_norm_weight_to_fp16)[name = string("last_cast_fp16")];
            tensor<fp16, [151936]> linear_196_bias_0_to_fp16 = const()[name = string("linear_196_bias_0_to_fp16"), val = tensor<fp16, [151936]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752502080)))];
            tensor<fp16, [1, 1, 151936]> linear_196_cast_fp16 = linear(bias = linear_196_bias_0_to_fp16, weight = model_embed_tokens_weight_to_fp16, x = last_cast_fp16)[name = string("linear_196_cast_fp16")];
            int32 var_9815_axis_0 = const()[name = string("op_9815_axis_0"), val = int32(-1)];
            bool var_9815_keep_dims_0 = const()[name = string("op_9815_keep_dims_0"), val = bool(false)];
            string var_9815_output_dtype_0 = const()[name = string("op_9815_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_9815_axis_0, keep_dims = var_9815_keep_dims_0, output_dtype = var_9815_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_9815_cast_fp16")];
            int32 var_9822 = const()[name = string("op_9822"), val = int32(0)];
            bool var_9823_interleave_0 = const()[name = string("op_9823_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_9822, interleave = var_9823_interleave_0, values = (k_1_cast_fp16, k_3_cast_fp16, k_5_cast_fp16, k_7_cast_fp16, k_9_cast_fp16, k_11_cast_fp16, k_13_cast_fp16, k_15_cast_fp16, k_17_cast_fp16, k_19_cast_fp16, k_21_cast_fp16, k_23_cast_fp16, k_25_cast_fp16, k_27_cast_fp16, k_29_cast_fp16, k_31_cast_fp16, k_33_cast_fp16, k_35_cast_fp16, k_37_cast_fp16, k_39_cast_fp16, k_41_cast_fp16, k_43_cast_fp16, k_45_cast_fp16, k_47_cast_fp16, k_49_cast_fp16, k_51_cast_fp16, k_53_cast_fp16, k_cast_fp16))[name = string("op_9823_cast_fp16")];
            int32 var_9825 = const()[name = string("op_9825"), val = int32(0)];
            bool var_9826_interleave_0 = const()[name = string("op_9826_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_9825, interleave = var_9826_interleave_0, values = (v_1_cast_fp16, v_3_cast_fp16, v_5_cast_fp16, v_7_cast_fp16, v_9_cast_fp16, v_11_cast_fp16, v_13_cast_fp16, v_15_cast_fp16, v_17_cast_fp16, v_19_cast_fp16, v_21_cast_fp16, v_23_cast_fp16, v_25_cast_fp16, v_27_cast_fp16, v_29_cast_fp16, v_31_cast_fp16, v_33_cast_fp16, v_35_cast_fp16, v_37_cast_fp16, v_39_cast_fp16, v_41_cast_fp16, v_43_cast_fp16, v_45_cast_fp16, v_47_cast_fp16, v_49_cast_fp16, v_51_cast_fp16, v_53_cast_fp16, v_cast_fp16))[name = string("op_9826_cast_fp16")];
        } -> (nextToken, newKeys, newValues);
    func window512<ios18>(tensor<int32, [1]> answerRow, tensor<fp16, [1, 1, 1, 512]> cacheMask, tensor<int32, [1, 16]> inputIds, state<tensor<fp16, [28, 8, 528, 128]>> keyCache, tensor<int32, [16]> positions, state<tensor<fp16, [28, 8, 528, 128]>> valueCache) {
            int32 x_1_batch_dims_0 = const()[name = string("x_1_batch_dims_0"), val = int32(0)];
            bool x_1_validate_indices_0 = const()[name = string("x_1_validate_indices_0"), val = bool(false)];
            tensor<fp16, [151936, 1024]> model_embed_tokens_weight_to_fp16 = const()[name = string("model_embed_tokens_weight_to_fp16"), val = tensor<fp16, [151936, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            int32 greater_equal_0_y_0 = const()[name = string("greater_equal_0_y_0"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0 = greater_equal(x = inputIds, y = greater_equal_0_y_0)[name = string("greater_equal_0")];
            int32 slice_by_index_0 = const()[name = string("slice_by_index_0"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0 = add(x = inputIds, y = slice_by_index_0)[name = string("add_0")];
            tensor<int32, [1, 16]> select_0 = select(a = inputIds, b = add_0, cond = greater_equal_0)[name = string("select_0")];
            int32 greater_equal_0_y_0_1 = const()[name = string("greater_equal_0_y_0_1"), val = int32(0)];
            tensor<bool, [1, 16]> greater_equal_0_1 = greater_equal(x = select_0, y = greater_equal_0_y_0_1)[name = string("greater_equal_0_1")];
            int32 slice_by_index_0_1 = const()[name = string("slice_by_index_0_1"), val = int32(151936)];
            tensor<int32, [1, 16]> add_0_1 = add(x = select_0, y = slice_by_index_0_1)[name = string("add_0_1")];
            tensor<int32, [1, 16]> select_0_1 = select(a = select_0, b = add_0_1, cond = greater_equal_0_1)[name = string("select_0_1")];
            int32 x_1_cast_fp16_axis_0 = const()[name = string("x_1_cast_fp16_axis_0"), val = int32(0)];
            tensor<fp16, [1, 16, 1024]> x_1_cast_fp16 = gather(axis = x_1_cast_fp16_axis_0, batch_dims = x_1_batch_dims_0, indices = select_0_1, validate_indices = x_1_validate_indices_0, x = model_embed_tokens_weight_to_fp16)[name = string("x_1_cast_fp16")];
            int32 var_1504 = const()[name = string("op_1504"), val = int32(0)];
            int32 var_1505_batch_dims_0 = const()[name = string("op_1505_batch_dims_0"), val = int32(0)];
            bool var_1505_validate_indices_0 = const()[name = string("op_1505_validate_indices_0"), val = bool(false)];
            tensor<fp16, [528, 64]> cos_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [528, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311165056))), scale = tensor<fp16, [528, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311198912))))[name = string("cos_to_fp16_quantized")];
            string positions_to_uint16_dtype_0 = const()[name = string("positions_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [16]> positions_to_uint16 = cast(dtype = positions_to_uint16_dtype_0, x = positions)[name = string("cast_1")];
            tensor<fp16, [16, 64]> var_1505_cast_fp16_cast_uint16 = gather(axis = var_1504, batch_dims = var_1505_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1505_validate_indices_0, x = cos_to_fp16_quantized)[name = string("op_1505_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1510 = const()[name = string("op_1510"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> cos_1_cast_fp16 = reshape(shape = var_1510, x = var_1505_cast_fp16_cast_uint16)[name = string("cos_1_cast_fp16")];
            int32 var_1512 = const()[name = string("op_1512"), val = int32(0)];
            int32 var_1513_batch_dims_0 = const()[name = string("op_1513_batch_dims_0"), val = int32(0)];
            bool var_1513_validate_indices_0 = const()[name = string("op_1513_validate_indices_0"), val = bool(false)];
            tensor<fp16, [528, 64]> sin_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [528, 64]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311200064))), scale = tensor<fp16, [528, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311233920))))[name = string("sin_to_fp16_quantized")];
            tensor<fp16, [16, 64]> var_1513_cast_fp16_cast_uint16 = gather(axis = var_1512, batch_dims = var_1513_batch_dims_0, indices = positions_to_uint16, validate_indices = var_1513_validate_indices_0, x = sin_to_fp16_quantized)[name = string("op_1513_cast_fp16_cast_uint16")];
            tensor<int32, [4]> var_1518 = const()[name = string("op_1518"), val = tensor<int32, [4]>([1, 1, 16, 64])];
            tensor<fp16, [1, 1, 16, 64]> sin_1_cast_fp16 = reshape(shape = var_1518, x = var_1513_cast_fp16_cast_uint16)[name = string("sin_1_cast_fp16")];
            fp16 var_1539_promoted_to_fp16 = const()[name = string("op_1539_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1541_cast_fp16 = pow(x = x_1_cast_fp16, y = var_1539_promoted_to_fp16)[name = string("op_1541_cast_fp16")];
            tensor<int32, [1]> var_1543_axes_0 = const()[name = string("op_1543_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1543_keep_dims_0 = const()[name = string("op_1543_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1543_cast_fp16 = reduce_mean(axes = var_1543_axes_0, keep_dims = var_1543_keep_dims_0, x = var_1541_cast_fp16)[name = string("op_1543_cast_fp16")];
            fp16 var_1544_to_fp16 = const()[name = string("op_1544_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1545_cast_fp16 = add(x = var_1543_cast_fp16, y = var_1544_to_fp16)[name = string("op_1545_cast_fp16")];
            fp32 var_1546_epsilon_0 = const()[name = string("op_1546_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1546_cast_fp16 = rsqrt(epsilon = var_1546_epsilon_0, x = var_1545_cast_fp16)[name = string("op_1546_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1547_cast_fp16 = mul(x = x_1_cast_fp16, y = var_1546_cast_fp16)[name = string("op_1547_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311235072)))];
            tensor<fp16, [1, 16, 1024]> input_1_cast_fp16 = mul(x = var_1547_cast_fp16, y = model_layers_0_input_layernorm_weight_to_fp16)[name = string("input_1_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_0_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(311237184))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313334400))))[name = string("model_layers_0_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [2048]> linear_0_bias_0_to_fp16 = const()[name = string("linear_0_bias_0_to_fp16"), val = tensor<fp16, [2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313338560)))];
            tensor<fp16, [1, 16, 2048]> linear_0_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_0_self_attn_q_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_0_cast_fp16")];
            tensor<int32, [4]> var_1556 = const()[name = string("op_1556"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1557_cast_fp16 = reshape(shape = var_1556, x = linear_0_cast_fp16)[name = string("op_1557_cast_fp16")];
            tensor<int32, [4]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1565_promoted_to_fp16 = const()[name = string("op_1565_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_1557_cast_fp16)[name = string("transpose_139")];
            tensor<fp16, [1, 16, 16, 128]> var_1567_cast_fp16 = pow(x = x_3_cast_fp16, y = var_1565_promoted_to_fp16)[name = string("op_1567_cast_fp16")];
            tensor<int32, [1]> var_1569_axes_0 = const()[name = string("op_1569_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1569_keep_dims_0 = const()[name = string("op_1569_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1569_cast_fp16 = reduce_mean(axes = var_1569_axes_0, keep_dims = var_1569_keep_dims_0, x = var_1567_cast_fp16)[name = string("op_1569_cast_fp16")];
            fp16 var_1570_to_fp16 = const()[name = string("op_1570_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1571_cast_fp16 = add(x = var_1569_cast_fp16, y = var_1570_to_fp16)[name = string("op_1571_cast_fp16")];
            fp32 var_1572_epsilon_0 = const()[name = string("op_1572_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1572_cast_fp16 = rsqrt(epsilon = var_1572_epsilon_0, x = var_1571_cast_fp16)[name = string("op_1572_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1573_cast_fp16 = mul(x = x_3_cast_fp16, y = var_1572_cast_fp16)[name = string("op_1573_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313342720)))];
            tensor<fp16, [1, 16, 16, 128]> x_7_cast_fp16 = mul(x = var_1573_cast_fp16, y = model_layers_0_self_attn_q_norm_weight_to_fp16)[name = string("x_7_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(313343040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314391680))))[name = string("model_layers_0_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1024]> linear_1_bias_0_to_fp16 = const()[name = string("linear_1_bias_0_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314393792)))];
            tensor<fp16, [1, 16, 1024]> linear_1_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_k_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_1_cast_fp16")];
            tensor<int32, [4]> var_1582 = const()[name = string("op_1582"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1583_cast_fp16 = reshape(shape = var_1582, x = linear_1_cast_fp16)[name = string("op_1583_cast_fp16")];
            tensor<int32, [4]> x_5_perm_0 = const()[name = string("x_5_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1591_promoted_to_fp16 = const()[name = string("op_1591_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_5_cast_fp16 = transpose(perm = x_5_perm_0, x = var_1583_cast_fp16)[name = string("transpose_138")];
            tensor<fp16, [1, 8, 16, 128]> var_1593_cast_fp16 = pow(x = x_5_cast_fp16, y = var_1591_promoted_to_fp16)[name = string("op_1593_cast_fp16")];
            tensor<int32, [1]> var_1595_axes_0 = const()[name = string("op_1595_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1595_keep_dims_0 = const()[name = string("op_1595_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1595_cast_fp16 = reduce_mean(axes = var_1595_axes_0, keep_dims = var_1595_keep_dims_0, x = var_1593_cast_fp16)[name = string("op_1595_cast_fp16")];
            fp16 var_1596_to_fp16 = const()[name = string("op_1596_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1597_cast_fp16 = add(x = var_1595_cast_fp16, y = var_1596_to_fp16)[name = string("op_1597_cast_fp16")];
            fp32 var_1598_epsilon_0 = const()[name = string("op_1598_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1598_cast_fp16 = rsqrt(epsilon = var_1598_epsilon_0, x = var_1597_cast_fp16)[name = string("op_1598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1599_cast_fp16 = mul(x = x_5_cast_fp16, y = var_1598_cast_fp16)[name = string("op_1599_cast_fp16")];
            tensor<fp16, [128]> model_layers_0_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_0_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314395904)))];
            tensor<fp16, [1, 8, 16, 128]> x_9_cast_fp16 = mul(x = var_1599_cast_fp16, y = model_layers_0_self_attn_k_norm_weight_to_fp16)[name = string("x_9_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_0_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(314396224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315444864))))[name = string("model_layers_0_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_2_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_v_proj_weight_to_fp16_quantized, x = input_1_cast_fp16)[name = string("linear_2_cast_fp16")];
            tensor<int32, [4]> var_1608 = const()[name = string("op_1608"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1609_cast_fp16 = reshape(shape = var_1608, x = linear_2_cast_fp16)[name = string("op_1609_cast_fp16")];
            tensor<int32, [4]> v_1_perm_0 = const()[name = string("v_1_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_1_begin_0 = const()[name = string("first_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_1_end_0 = const()[name = string("first_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_1_end_mask_0 = const()[name = string("first_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_1_cast_fp16 = slice_by_index(begin = first_1_begin_0, end = first_1_end_0, end_mask = first_1_end_mask_0, x = x_7_cast_fp16)[name = string("first_1_cast_fp16")];
            tensor<int32, [4]> second_1_begin_0 = const()[name = string("second_1_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_1_end_0 = const()[name = string("second_1_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_1_end_mask_0 = const()[name = string("second_1_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_1_cast_fp16 = slice_by_index(begin = second_1_begin_0, end = second_1_end_0, end_mask = second_1_end_mask_0, x = x_7_cast_fp16)[name = string("second_1_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1623_cast_fp16 = mul(x = first_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1623_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1624_cast_fp16 = mul(x = second_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1624_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1626_cast_fp16 = sub(x = var_1623_cast_fp16, y = var_1624_cast_fp16)[name = string("op_1626_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1627_cast_fp16 = mul(x = second_1_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1627_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1628_cast_fp16 = mul(x = first_1_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1628_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1630_cast_fp16 = add(x = var_1627_cast_fp16, y = var_1628_cast_fp16)[name = string("op_1630_cast_fp16")];
            int32 var_1632 = const()[name = string("op_1632"), val = int32(-1)];
            bool q_1_interleave_0 = const()[name = string("q_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_1_cast_fp16 = concat(axis = var_1632, interleave = q_1_interleave_0, values = (var_1626_cast_fp16, var_1630_cast_fp16))[name = string("q_1_cast_fp16")];
            tensor<int32, [4]> first_3_begin_0 = const()[name = string("first_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_3_end_0 = const()[name = string("first_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_3_end_mask_0 = const()[name = string("first_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_3_cast_fp16 = slice_by_index(begin = first_3_begin_0, end = first_3_end_0, end_mask = first_3_end_mask_0, x = x_9_cast_fp16)[name = string("first_3_cast_fp16")];
            tensor<int32, [4]> second_3_begin_0 = const()[name = string("second_3_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_3_end_0 = const()[name = string("second_3_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_3_end_mask_0 = const()[name = string("second_3_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_3_cast_fp16 = slice_by_index(begin = second_3_begin_0, end = second_3_end_0, end_mask = second_3_end_mask_0, x = x_9_cast_fp16)[name = string("second_3_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1644_cast_fp16 = mul(x = first_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1644_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1645_cast_fp16 = mul(x = second_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1645_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1647_cast_fp16 = sub(x = var_1644_cast_fp16, y = var_1645_cast_fp16)[name = string("op_1647_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1648_cast_fp16 = mul(x = second_3_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1648_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1649_cast_fp16 = mul(x = first_3_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1649_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1651_cast_fp16 = add(x = var_1648_cast_fp16, y = var_1649_cast_fp16)[name = string("op_1651_cast_fp16")];
            int32 var_1653 = const()[name = string("op_1653"), val = int32(-1)];
            bool k_1_interleave_0 = const()[name = string("k_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_1_cast_fp16 = concat(axis = var_1653, interleave = k_1_interleave_0, values = (var_1647_cast_fp16, var_1651_cast_fp16))[name = string("k_1_cast_fp16")];
            tensor<int32, [4]> var_1664_begin_0 = const()[name = string("op_1664_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1664_end_0 = const()[name = string("op_1664_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1664_end_mask_0 = const()[name = string("op_1664_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1664_stride_0 = const()[name = string("op_1664_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1664_cast_fp16 = slice_by_index(begin = var_1664_begin_0, end = var_1664_end_0, end_mask = var_1664_end_mask_0, stride = var_1664_stride_0, x = q_1_cast_fp16)[name = string("op_1664_cast_fp16")];
            tensor<int32, [4]> var_1674_begin_0 = const()[name = string("op_1674_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1674_end_0 = const()[name = string("op_1674_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1674_end_mask_0 = const()[name = string("op_1674_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1674_stride_0 = const()[name = string("op_1674_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1674_cast_fp16 = slice_by_index(begin = var_1674_begin_0, end = var_1674_end_0, end_mask = var_1674_end_mask_0, stride = var_1674_stride_0, x = q_1_cast_fp16)[name = string("op_1674_cast_fp16")];
            int32 var_1676 = const()[name = string("op_1676"), val = int32(2)];
            bool stacked_1_interleave_0 = const()[name = string("stacked_1_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_1_cast_fp16 = concat(axis = var_1676, interleave = stacked_1_interleave_0, values = (var_1664_cast_fp16, var_1674_cast_fp16))[name = string("stacked_1_cast_fp16")];
            tensor<fp16, [28, 8, 528, 128]> read_state_0 = read_state(input = keyCache)[name = string("read_state_0")];
            tensor<int32, [4]> var_1682_begin_0 = const()[name = string("op_1682_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1682_end_0 = const()[name = string("op_1682_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_1682_end_mask_0 = const()[name = string("op_1682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_1682_cast_fp16 = slice_by_index(begin = var_1682_begin_0, end = var_1682_end_0, end_mask = var_1682_end_mask_0, x = read_state_0)[name = string("op_1682_cast_fp16")];
            tensor<int32, [4]> var_1692_begin_0 = const()[name = string("op_1692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1692_end_0 = const()[name = string("op_1692_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_1692_end_mask_0 = const()[name = string("op_1692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_1692_cast_fp16 = slice_by_index(begin = var_1692_begin_0, end = var_1692_end_0, end_mask = var_1692_end_mask_0, x = var_1682_cast_fp16)[name = string("op_1692_cast_fp16")];
            bool var_1701_transpose_x_1 = const()[name = string("op_1701_transpose_x_1"), val = bool(false)];
            bool var_1701_transpose_y_1 = const()[name = string("op_1701_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_1701_cast_fp16 = matmul(transpose_x = var_1701_transpose_x_1, transpose_y = var_1701_transpose_y_1, x = var_1692_cast_fp16, y = stacked_1_cast_fp16)[name = string("op_1701_cast_fp16")];
            tensor<int32, [4]> var_1704_perm_0 = const()[name = string("op_1704_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_1705_to_fp16 = const()[name = string("op_1705_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_1704_cast_fp16 = transpose(perm = var_1704_perm_0, x = var_1701_cast_fp16)[name = string("transpose_137")];
            tensor<fp16, [1, 8, 32, 512]> var_1706_cast_fp16 = mul(x = var_1704_cast_fp16, y = var_1705_to_fp16)[name = string("op_1706_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_1_cast_fp16 = add(x = var_1706_cast_fp16, y = cacheMask)[name = string("cache_weights_1_cast_fp16")];
            bool var_1712_transpose_x_1 = const()[name = string("op_1712_transpose_x_1"), val = bool(false)];
            bool var_1712_transpose_y_1 = const()[name = string("op_1712_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_1712_cast_fp16 = matmul(transpose_x = var_1712_transpose_x_1, transpose_y = var_1712_transpose_y_1, x = stacked_1_cast_fp16, y = k_1_cast_fp16)[name = string("op_1712_cast_fp16")];
            fp16 var_1713_to_fp16 = const()[name = string("op_1713_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_1714_cast_fp16 = mul(x = var_1712_cast_fp16, y = var_1713_to_fp16)[name = string("op_1714_cast_fp16")];
            tensor<fp16, [32, 16]> later_rows_mask_to_fp16 = const()[name = string("later_rows_mask_to_fp16"), val = tensor<fp16, [32, 16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315446976)))];
            tensor<fp16, [1, 8, 32, 16]> call_weights_1_cast_fp16 = add(x = var_1714_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_1_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, 8, 32, 528]> var_1719_cast_fp16 = concat(axis = var_1718, interleave = var_1719_interleave_0, values = (cache_weights_1_cast_fp16, call_weights_1_cast_fp16))[name = string("op_1719_cast_fp16")];
            int32 var_1720 = const()[name = string("op_1720"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_1_cast_fp16 = softmax(axis = var_1720, x = var_1719_cast_fp16)[name = string("shares_1_cast_fp16")];
            tensor<int32, [4]> var_1727_begin_0 = const()[name = string("op_1727_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1727_end_0 = const()[name = string("op_1727_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_1727_end_mask_0 = const()[name = string("op_1727_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_1727_cast_fp16 = slice_by_index(begin = var_1727_begin_0, end = var_1727_end_0, end_mask = var_1727_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1727_cast_fp16")];
            tensor<fp16, [28, 8, 528, 128]> read_state_1 = read_state(input = valueCache)[name = string("read_state_1")];
            tensor<int32, [4]> var_1732_begin_0 = const()[name = string("op_1732_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1732_end_0 = const()[name = string("op_1732_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_1732_end_mask_0 = const()[name = string("op_1732_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_1732_cast_fp16 = slice_by_index(begin = var_1732_begin_0, end = var_1732_end_0, end_mask = var_1732_end_mask_0, x = read_state_1)[name = string("op_1732_cast_fp16")];
            tensor<int32, [4]> var_1742_begin_0 = const()[name = string("op_1742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1742_end_0 = const()[name = string("op_1742_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_1742_end_mask_0 = const()[name = string("op_1742_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_1742_cast_fp16 = slice_by_index(begin = var_1742_begin_0, end = var_1742_end_0, end_mask = var_1742_end_mask_0, x = var_1732_cast_fp16)[name = string("op_1742_cast_fp16")];
            bool var_1748_transpose_x_0 = const()[name = string("op_1748_transpose_x_0"), val = bool(false)];
            bool var_1748_transpose_y_0 = const()[name = string("op_1748_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_1748_cast_fp16 = matmul(transpose_x = var_1748_transpose_x_0, transpose_y = var_1748_transpose_y_0, x = var_1727_cast_fp16, y = var_1742_cast_fp16)[name = string("op_1748_cast_fp16")];
            tensor<int32, [4]> var_1753_begin_0 = const()[name = string("op_1753_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_1753_end_0 = const()[name = string("op_1753_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_1753_end_mask_0 = const()[name = string("op_1753_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_1753_cast_fp16 = slice_by_index(begin = var_1753_begin_0, end = var_1753_end_0, end_mask = var_1753_end_mask_0, x = shares_1_cast_fp16)[name = string("op_1753_cast_fp16")];
            bool var_1754_transpose_x_0 = const()[name = string("op_1754_transpose_x_0"), val = bool(false)];
            bool var_1754_transpose_y_0 = const()[name = string("op_1754_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_1_cast_fp16 = transpose(perm = v_1_perm_0, x = var_1609_cast_fp16)[name = string("transpose_136")];
            tensor<fp16, [1, 8, 32, 128]> var_1754_cast_fp16 = matmul(transpose_x = var_1754_transpose_x_0, transpose_y = var_1754_transpose_y_0, x = var_1753_cast_fp16, y = v_1_cast_fp16)[name = string("op_1754_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_1_cast_fp16 = add(x = var_1748_cast_fp16, y = var_1754_cast_fp16)[name = string("mixed_1_cast_fp16")];
            tensor<int32, [4]> var_1771_begin_0 = const()[name = string("op_1771_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1771_end_0 = const()[name = string("op_1771_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_1771_end_mask_0 = const()[name = string("op_1771_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1771_cast_fp16 = slice_by_index(begin = var_1771_begin_0, end = var_1771_end_0, end_mask = var_1771_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1771_cast_fp16")];
            tensor<int32, [4]> var_1786_begin_0 = const()[name = string("op_1786_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_1786_end_0 = const()[name = string("op_1786_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_1786_end_mask_0 = const()[name = string("op_1786_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_1786_cast_fp16 = slice_by_index(begin = var_1786_begin_0, end = var_1786_end_0, end_mask = var_1786_end_mask_0, x = mixed_1_cast_fp16)[name = string("op_1786_cast_fp16")];
            int32 var_1788 = const()[name = string("op_1788"), val = int32(1)];
            bool mixed_3_interleave_0 = const()[name = string("mixed_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_3_cast_fp16 = concat(axis = var_1788, interleave = mixed_3_interleave_0, values = (var_1771_cast_fp16, var_1786_cast_fp16))[name = string("mixed_3_cast_fp16")];
            tensor<int32, [4]> var_1792_perm_0 = const()[name = string("op_1792_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_1796 = const()[name = string("op_1796"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_1792_cast_fp16 = transpose(perm = var_1792_perm_0, x = mixed_3_cast_fp16)[name = string("transpose_135")];
            tensor<fp16, [1, 16, 2048]> input_3_cast_fp16 = reshape(shape = var_1796, x = var_1792_cast_fp16)[name = string("input_3_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_0_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(315448064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317545280))))[name = string("model_layers_0_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_3_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_self_attn_o_proj_weight_to_fp16_quantized, x = input_3_cast_fp16)[name = string("linear_3_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_11_cast_fp16 = add(x = x_1_cast_fp16, y = linear_3_cast_fp16)[name = string("x_11_cast_fp16")];
            fp16 var_1807_promoted_to_fp16 = const()[name = string("op_1807_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1809_cast_fp16 = pow(x = x_11_cast_fp16, y = var_1807_promoted_to_fp16)[name = string("op_1809_cast_fp16")];
            tensor<int32, [1]> var_1811_axes_0 = const()[name = string("op_1811_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1811_keep_dims_0 = const()[name = string("op_1811_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1811_cast_fp16 = reduce_mean(axes = var_1811_axes_0, keep_dims = var_1811_keep_dims_0, x = var_1809_cast_fp16)[name = string("op_1811_cast_fp16")];
            fp16 var_1812_to_fp16 = const()[name = string("op_1812_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1813_cast_fp16 = add(x = var_1811_cast_fp16, y = var_1812_to_fp16)[name = string("op_1813_cast_fp16")];
            fp32 var_1814_epsilon_0 = const()[name = string("op_1814_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1814_cast_fp16 = rsqrt(epsilon = var_1814_epsilon_0, x = var_1813_cast_fp16)[name = string("op_1814_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1815_cast_fp16 = mul(x = x_11_cast_fp16, y = var_1814_cast_fp16)[name = string("op_1815_cast_fp16")];
            tensor<fp16, [1024]> model_layers_0_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_0_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317547392)))];
            tensor<fp16, [1, 16, 1024]> input_5_cast_fp16 = mul(x = var_1815_cast_fp16, y = model_layers_0_post_attention_layernorm_weight_to_fp16)[name = string("input_5_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(317549504))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320695296))))[name = string("model_layers_0_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [3072]> linear_4_bias_0_to_fp16 = const()[name = string("linear_4_bias_0_to_fp16"), val = tensor<fp16, [3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320701504)))];
            tensor<fp16, [1, 16, 3072]> linear_4_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_gate_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_4_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_1820_cast_fp16 = silu(x = linear_4_cast_fp16)[name = string("op_1820_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_0_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(320707712))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323853504))))[name = string("model_layers_0_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_5_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_0_mlp_up_proj_weight_to_fp16_quantized, x = input_5_cast_fp16)[name = string("linear_5_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_9_cast_fp16 = mul(x = var_1820_cast_fp16, y = linear_5_cast_fp16)[name = string("input_9_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_0_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(323859712))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327005504))))[name = string("model_layers_0_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_6_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_0_mlp_down_proj_weight_to_fp16_quantized, x = input_9_cast_fp16)[name = string("linear_6_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_13_cast_fp16 = add(x = x_11_cast_fp16, y = linear_6_cast_fp16)[name = string("x_13_cast_fp16")];
            fp16 var_1834_promoted_to_fp16 = const()[name = string("op_1834_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_1836_cast_fp16 = pow(x = x_13_cast_fp16, y = var_1834_promoted_to_fp16)[name = string("op_1836_cast_fp16")];
            tensor<int32, [1]> var_1838_axes_0 = const()[name = string("op_1838_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1838_keep_dims_0 = const()[name = string("op_1838_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_1838_cast_fp16 = reduce_mean(axes = var_1838_axes_0, keep_dims = var_1838_keep_dims_0, x = var_1836_cast_fp16)[name = string("op_1838_cast_fp16")];
            fp16 var_1839_to_fp16 = const()[name = string("op_1839_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_1840_cast_fp16 = add(x = var_1838_cast_fp16, y = var_1839_to_fp16)[name = string("op_1840_cast_fp16")];
            fp32 var_1841_epsilon_0 = const()[name = string("op_1841_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_1841_cast_fp16 = rsqrt(epsilon = var_1841_epsilon_0, x = var_1840_cast_fp16)[name = string("op_1841_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_1842_cast_fp16 = mul(x = x_13_cast_fp16, y = var_1841_cast_fp16)[name = string("op_1842_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327007616)))];
            tensor<fp16, [1, 16, 1024]> input_11_cast_fp16 = mul(x = var_1842_cast_fp16, y = model_layers_1_input_layernorm_weight_to_fp16)[name = string("input_11_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_1_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(327009728))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329106944))))[name = string("model_layers_1_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_7_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_1_self_attn_q_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_7_cast_fp16")];
            tensor<int32, [4]> var_1851 = const()[name = string("op_1851"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_1852_cast_fp16 = reshape(shape = var_1851, x = linear_7_cast_fp16)[name = string("op_1852_cast_fp16")];
            tensor<int32, [4]> x_15_perm_0 = const()[name = string("x_15_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1860_promoted_to_fp16 = const()[name = string("op_1860_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_15_cast_fp16 = transpose(perm = x_15_perm_0, x = var_1852_cast_fp16)[name = string("transpose_134")];
            tensor<fp16, [1, 16, 16, 128]> var_1862_cast_fp16 = pow(x = x_15_cast_fp16, y = var_1860_promoted_to_fp16)[name = string("op_1862_cast_fp16")];
            tensor<int32, [1]> var_1864_axes_0 = const()[name = string("op_1864_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1864_keep_dims_0 = const()[name = string("op_1864_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_1864_cast_fp16 = reduce_mean(axes = var_1864_axes_0, keep_dims = var_1864_keep_dims_0, x = var_1862_cast_fp16)[name = string("op_1864_cast_fp16")];
            fp16 var_1865_to_fp16 = const()[name = string("op_1865_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_1866_cast_fp16 = add(x = var_1864_cast_fp16, y = var_1865_to_fp16)[name = string("op_1866_cast_fp16")];
            fp32 var_1867_epsilon_0 = const()[name = string("op_1867_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_1867_cast_fp16 = rsqrt(epsilon = var_1867_epsilon_0, x = var_1866_cast_fp16)[name = string("op_1867_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_1868_cast_fp16 = mul(x = x_15_cast_fp16, y = var_1867_cast_fp16)[name = string("op_1868_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111104)))];
            tensor<fp16, [1, 16, 16, 128]> x_19_cast_fp16 = mul(x = var_1868_cast_fp16, y = model_layers_1_self_attn_q_norm_weight_to_fp16)[name = string("x_19_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(329111424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330160064))))[name = string("model_layers_1_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_8_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_k_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_8_cast_fp16")];
            tensor<int32, [4]> var_1877 = const()[name = string("op_1877"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1878_cast_fp16 = reshape(shape = var_1877, x = linear_8_cast_fp16)[name = string("op_1878_cast_fp16")];
            tensor<int32, [4]> x_17_perm_0 = const()[name = string("x_17_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_1886_promoted_to_fp16 = const()[name = string("op_1886_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_17_cast_fp16 = transpose(perm = x_17_perm_0, x = var_1878_cast_fp16)[name = string("transpose_133")];
            tensor<fp16, [1, 8, 16, 128]> var_1888_cast_fp16 = pow(x = x_17_cast_fp16, y = var_1886_promoted_to_fp16)[name = string("op_1888_cast_fp16")];
            tensor<int32, [1]> var_1890_axes_0 = const()[name = string("op_1890_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_1890_keep_dims_0 = const()[name = string("op_1890_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_1890_cast_fp16 = reduce_mean(axes = var_1890_axes_0, keep_dims = var_1890_keep_dims_0, x = var_1888_cast_fp16)[name = string("op_1890_cast_fp16")];
            fp16 var_1891_to_fp16 = const()[name = string("op_1891_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_1892_cast_fp16 = add(x = var_1890_cast_fp16, y = var_1891_to_fp16)[name = string("op_1892_cast_fp16")];
            fp32 var_1893_epsilon_0 = const()[name = string("op_1893_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_1893_cast_fp16 = rsqrt(epsilon = var_1893_epsilon_0, x = var_1892_cast_fp16)[name = string("op_1893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_1894_cast_fp16 = mul(x = x_17_cast_fp16, y = var_1893_cast_fp16)[name = string("op_1894_cast_fp16")];
            tensor<fp16, [128]> model_layers_1_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_1_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162176)))];
            tensor<fp16, [1, 8, 16, 128]> x_21_cast_fp16 = mul(x = var_1894_cast_fp16, y = model_layers_1_self_attn_k_norm_weight_to_fp16)[name = string("x_21_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_1_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(330162496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331211136))))[name = string("model_layers_1_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_9_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_v_proj_weight_to_fp16_quantized, x = input_11_cast_fp16)[name = string("linear_9_cast_fp16")];
            tensor<int32, [4]> var_1903 = const()[name = string("op_1903"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_1904_cast_fp16 = reshape(shape = var_1903, x = linear_9_cast_fp16)[name = string("op_1904_cast_fp16")];
            tensor<int32, [4]> v_3_perm_0 = const()[name = string("v_3_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_5_begin_0 = const()[name = string("first_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_5_end_0 = const()[name = string("first_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_5_end_mask_0 = const()[name = string("first_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_5_cast_fp16 = slice_by_index(begin = first_5_begin_0, end = first_5_end_0, end_mask = first_5_end_mask_0, x = x_19_cast_fp16)[name = string("first_5_cast_fp16")];
            tensor<int32, [4]> second_5_begin_0 = const()[name = string("second_5_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_5_end_0 = const()[name = string("second_5_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_5_end_mask_0 = const()[name = string("second_5_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_5_cast_fp16 = slice_by_index(begin = second_5_begin_0, end = second_5_end_0, end_mask = second_5_end_mask_0, x = x_19_cast_fp16)[name = string("second_5_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1918_cast_fp16 = mul(x = first_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1918_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1919_cast_fp16 = mul(x = second_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1919_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1921_cast_fp16 = sub(x = var_1918_cast_fp16, y = var_1919_cast_fp16)[name = string("op_1921_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1922_cast_fp16 = mul(x = second_5_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1922_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1923_cast_fp16 = mul(x = first_5_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1923_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_1925_cast_fp16 = add(x = var_1922_cast_fp16, y = var_1923_cast_fp16)[name = string("op_1925_cast_fp16")];
            int32 var_1927 = const()[name = string("op_1927"), val = int32(-1)];
            bool q_3_interleave_0 = const()[name = string("q_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_3_cast_fp16 = concat(axis = var_1927, interleave = q_3_interleave_0, values = (var_1921_cast_fp16, var_1925_cast_fp16))[name = string("q_3_cast_fp16")];
            tensor<int32, [4]> first_7_begin_0 = const()[name = string("first_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_7_end_0 = const()[name = string("first_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_7_end_mask_0 = const()[name = string("first_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_7_cast_fp16 = slice_by_index(begin = first_7_begin_0, end = first_7_end_0, end_mask = first_7_end_mask_0, x = x_21_cast_fp16)[name = string("first_7_cast_fp16")];
            tensor<int32, [4]> second_7_begin_0 = const()[name = string("second_7_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_7_end_0 = const()[name = string("second_7_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_7_end_mask_0 = const()[name = string("second_7_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_7_cast_fp16 = slice_by_index(begin = second_7_begin_0, end = second_7_end_0, end_mask = second_7_end_mask_0, x = x_21_cast_fp16)[name = string("second_7_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1939_cast_fp16 = mul(x = first_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1939_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1940_cast_fp16 = mul(x = second_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1940_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1942_cast_fp16 = sub(x = var_1939_cast_fp16, y = var_1940_cast_fp16)[name = string("op_1942_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1943_cast_fp16 = mul(x = second_7_cast_fp16, y = cos_1_cast_fp16)[name = string("op_1943_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1944_cast_fp16 = mul(x = first_7_cast_fp16, y = sin_1_cast_fp16)[name = string("op_1944_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_1946_cast_fp16 = add(x = var_1943_cast_fp16, y = var_1944_cast_fp16)[name = string("op_1946_cast_fp16")];
            int32 var_1948 = const()[name = string("op_1948"), val = int32(-1)];
            bool k_3_interleave_0 = const()[name = string("k_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_3_cast_fp16 = concat(axis = var_1948, interleave = k_3_interleave_0, values = (var_1942_cast_fp16, var_1946_cast_fp16))[name = string("k_3_cast_fp16")];
            tensor<int32, [4]> var_1959_begin_0 = const()[name = string("op_1959_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1959_end_0 = const()[name = string("op_1959_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1959_end_mask_0 = const()[name = string("op_1959_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1959_stride_0 = const()[name = string("op_1959_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1959_cast_fp16 = slice_by_index(begin = var_1959_begin_0, end = var_1959_end_0, end_mask = var_1959_end_mask_0, stride = var_1959_stride_0, x = q_3_cast_fp16)[name = string("op_1959_cast_fp16")];
            tensor<int32, [4]> var_1969_begin_0 = const()[name = string("op_1969_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_1969_end_0 = const()[name = string("op_1969_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_1969_end_mask_0 = const()[name = string("op_1969_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_1969_stride_0 = const()[name = string("op_1969_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_1969_cast_fp16 = slice_by_index(begin = var_1969_begin_0, end = var_1969_end_0, end_mask = var_1969_end_mask_0, stride = var_1969_stride_0, x = q_3_cast_fp16)[name = string("op_1969_cast_fp16")];
            int32 var_1971 = const()[name = string("op_1971"), val = int32(2)];
            bool stacked_3_interleave_0 = const()[name = string("stacked_3_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_3_cast_fp16 = concat(axis = var_1971, interleave = stacked_3_interleave_0, values = (var_1959_cast_fp16, var_1969_cast_fp16))[name = string("stacked_3_cast_fp16")];
            tensor<int32, [4]> var_1977_begin_0 = const()[name = string("op_1977_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_1977_end_0 = const()[name = string("op_1977_end_0"), val = tensor<int32, [4]>([2, 8, 528, 128])];
            tensor<bool, [4]> var_1977_end_mask_0 = const()[name = string("op_1977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_1977_cast_fp16 = slice_by_index(begin = var_1977_begin_0, end = var_1977_end_0, end_mask = var_1977_end_mask_0, x = read_state_0)[name = string("op_1977_cast_fp16")];
            tensor<int32, [4]> var_1987_begin_0 = const()[name = string("op_1987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_1987_end_0 = const()[name = string("op_1987_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_1987_end_mask_0 = const()[name = string("op_1987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_1987_cast_fp16 = slice_by_index(begin = var_1987_begin_0, end = var_1987_end_0, end_mask = var_1987_end_mask_0, x = var_1977_cast_fp16)[name = string("op_1987_cast_fp16")];
            bool var_1996_transpose_x_1 = const()[name = string("op_1996_transpose_x_1"), val = bool(false)];
            bool var_1996_transpose_y_1 = const()[name = string("op_1996_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_1996_cast_fp16 = matmul(transpose_x = var_1996_transpose_x_1, transpose_y = var_1996_transpose_y_1, x = var_1987_cast_fp16, y = stacked_3_cast_fp16)[name = string("op_1996_cast_fp16")];
            tensor<int32, [4]> var_1999_perm_0 = const()[name = string("op_1999_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2000_to_fp16 = const()[name = string("op_2000_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_1999_cast_fp16 = transpose(perm = var_1999_perm_0, x = var_1996_cast_fp16)[name = string("transpose_132")];
            tensor<fp16, [1, 8, 32, 512]> var_2001_cast_fp16 = mul(x = var_1999_cast_fp16, y = var_2000_to_fp16)[name = string("op_2001_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_3_cast_fp16 = add(x = var_2001_cast_fp16, y = cacheMask)[name = string("cache_weights_3_cast_fp16")];
            bool var_2007_transpose_x_1 = const()[name = string("op_2007_transpose_x_1"), val = bool(false)];
            bool var_2007_transpose_y_1 = const()[name = string("op_2007_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2007_cast_fp16 = matmul(transpose_x = var_2007_transpose_x_1, transpose_y = var_2007_transpose_y_1, x = stacked_3_cast_fp16, y = k_3_cast_fp16)[name = string("op_2007_cast_fp16")];
            fp16 var_2008_to_fp16 = const()[name = string("op_2008_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2009_cast_fp16 = mul(x = var_2007_cast_fp16, y = var_2008_to_fp16)[name = string("op_2009_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_3_cast_fp16 = add(x = var_2009_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_3_cast_fp16")];
            int32 var_2013 = const()[name = string("op_2013"), val = int32(-1)];
            bool var_2014_interleave_0 = const()[name = string("op_2014_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2014_cast_fp16 = concat(axis = var_2013, interleave = var_2014_interleave_0, values = (cache_weights_3_cast_fp16, call_weights_3_cast_fp16))[name = string("op_2014_cast_fp16")];
            int32 var_2015 = const()[name = string("op_2015"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_3_cast_fp16 = softmax(axis = var_2015, x = var_2014_cast_fp16)[name = string("shares_3_cast_fp16")];
            tensor<int32, [4]> var_2022_begin_0 = const()[name = string("op_2022_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2022_end_0 = const()[name = string("op_2022_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2022_end_mask_0 = const()[name = string("op_2022_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2022_cast_fp16 = slice_by_index(begin = var_2022_begin_0, end = var_2022_end_0, end_mask = var_2022_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2022_cast_fp16")];
            tensor<int32, [4]> var_2027_begin_0 = const()[name = string("op_2027_begin_0"), val = tensor<int32, [4]>([1, 0, 0, 0])];
            tensor<int32, [4]> var_2027_end_0 = const()[name = string("op_2027_end_0"), val = tensor<int32, [4]>([2, 8, 528, 128])];
            tensor<bool, [4]> var_2027_end_mask_0 = const()[name = string("op_2027_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2027_cast_fp16 = slice_by_index(begin = var_2027_begin_0, end = var_2027_end_0, end_mask = var_2027_end_mask_0, x = read_state_1)[name = string("op_2027_cast_fp16")];
            tensor<int32, [4]> var_2037_begin_0 = const()[name = string("op_2037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2037_end_0 = const()[name = string("op_2037_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2037_end_mask_0 = const()[name = string("op_2037_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2037_cast_fp16 = slice_by_index(begin = var_2037_begin_0, end = var_2037_end_0, end_mask = var_2037_end_mask_0, x = var_2027_cast_fp16)[name = string("op_2037_cast_fp16")];
            bool var_2043_transpose_x_0 = const()[name = string("op_2043_transpose_x_0"), val = bool(false)];
            bool var_2043_transpose_y_0 = const()[name = string("op_2043_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2043_cast_fp16 = matmul(transpose_x = var_2043_transpose_x_0, transpose_y = var_2043_transpose_y_0, x = var_2022_cast_fp16, y = var_2037_cast_fp16)[name = string("op_2043_cast_fp16")];
            tensor<int32, [4]> var_2048_begin_0 = const()[name = string("op_2048_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2048_end_0 = const()[name = string("op_2048_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2048_end_mask_0 = const()[name = string("op_2048_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2048_cast_fp16 = slice_by_index(begin = var_2048_begin_0, end = var_2048_end_0, end_mask = var_2048_end_mask_0, x = shares_3_cast_fp16)[name = string("op_2048_cast_fp16")];
            bool var_2049_transpose_x_0 = const()[name = string("op_2049_transpose_x_0"), val = bool(false)];
            bool var_2049_transpose_y_0 = const()[name = string("op_2049_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_3_cast_fp16 = transpose(perm = v_3_perm_0, x = var_1904_cast_fp16)[name = string("transpose_131")];
            tensor<fp16, [1, 8, 32, 128]> var_2049_cast_fp16 = matmul(transpose_x = var_2049_transpose_x_0, transpose_y = var_2049_transpose_y_0, x = var_2048_cast_fp16, y = v_3_cast_fp16)[name = string("op_2049_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_5_cast_fp16 = add(x = var_2043_cast_fp16, y = var_2049_cast_fp16)[name = string("mixed_5_cast_fp16")];
            tensor<int32, [4]> var_2066_begin_0 = const()[name = string("op_2066_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2066_end_0 = const()[name = string("op_2066_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2066_end_mask_0 = const()[name = string("op_2066_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2066_cast_fp16 = slice_by_index(begin = var_2066_begin_0, end = var_2066_end_0, end_mask = var_2066_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2066_cast_fp16")];
            tensor<int32, [4]> var_2081_begin_0 = const()[name = string("op_2081_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2081_end_0 = const()[name = string("op_2081_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2081_end_mask_0 = const()[name = string("op_2081_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2081_cast_fp16 = slice_by_index(begin = var_2081_begin_0, end = var_2081_end_0, end_mask = var_2081_end_mask_0, x = mixed_5_cast_fp16)[name = string("op_2081_cast_fp16")];
            int32 var_2083 = const()[name = string("op_2083"), val = int32(1)];
            bool mixed_7_interleave_0 = const()[name = string("mixed_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_7_cast_fp16 = concat(axis = var_2083, interleave = mixed_7_interleave_0, values = (var_2066_cast_fp16, var_2081_cast_fp16))[name = string("mixed_7_cast_fp16")];
            tensor<int32, [4]> var_2087_perm_0 = const()[name = string("op_2087_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2091 = const()[name = string("op_2091"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2087_cast_fp16 = transpose(perm = var_2087_perm_0, x = mixed_7_cast_fp16)[name = string("transpose_130")];
            tensor<fp16, [1, 16, 2048]> input_13_cast_fp16 = reshape(shape = var_2091, x = var_2087_cast_fp16)[name = string("input_13_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_1_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(331213248))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333310464))))[name = string("model_layers_1_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_10_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_self_attn_o_proj_weight_to_fp16_quantized, x = input_13_cast_fp16)[name = string("linear_10_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_23_cast_fp16 = add(x = x_13_cast_fp16, y = linear_10_cast_fp16)[name = string("x_23_cast_fp16")];
            fp16 var_2102_promoted_to_fp16 = const()[name = string("op_2102_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2104_cast_fp16 = pow(x = x_23_cast_fp16, y = var_2102_promoted_to_fp16)[name = string("op_2104_cast_fp16")];
            tensor<int32, [1]> var_2106_axes_0 = const()[name = string("op_2106_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2106_keep_dims_0 = const()[name = string("op_2106_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2106_cast_fp16 = reduce_mean(axes = var_2106_axes_0, keep_dims = var_2106_keep_dims_0, x = var_2104_cast_fp16)[name = string("op_2106_cast_fp16")];
            fp16 var_2107_to_fp16 = const()[name = string("op_2107_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2108_cast_fp16 = add(x = var_2106_cast_fp16, y = var_2107_to_fp16)[name = string("op_2108_cast_fp16")];
            fp32 var_2109_epsilon_0 = const()[name = string("op_2109_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2109_cast_fp16 = rsqrt(epsilon = var_2109_epsilon_0, x = var_2108_cast_fp16)[name = string("op_2109_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2110_cast_fp16 = mul(x = x_23_cast_fp16, y = var_2109_cast_fp16)[name = string("op_2110_cast_fp16")];
            tensor<fp16, [1024]> model_layers_1_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_1_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333312576)))];
            tensor<fp16, [1, 16, 1024]> input_15_cast_fp16 = mul(x = var_2110_cast_fp16, y = model_layers_1_post_attention_layernorm_weight_to_fp16)[name = string("input_15_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(333314688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336460480))))[name = string("model_layers_1_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_11_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_gate_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_11_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2115_cast_fp16 = silu(x = linear_11_cast_fp16)[name = string("op_2115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_1_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(336466688))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339612480))))[name = string("model_layers_1_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_12_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_1_mlp_up_proj_weight_to_fp16_quantized, x = input_15_cast_fp16)[name = string("linear_12_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_19_cast_fp16 = mul(x = var_2115_cast_fp16, y = linear_12_cast_fp16)[name = string("input_19_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_1_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(339618688))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342764480))))[name = string("model_layers_1_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_13_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_1_mlp_down_proj_weight_to_fp16_quantized, x = input_19_cast_fp16)[name = string("linear_13_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_25_cast_fp16 = add(x = x_23_cast_fp16, y = linear_13_cast_fp16)[name = string("x_25_cast_fp16")];
            fp16 var_2129_promoted_to_fp16 = const()[name = string("op_2129_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2131_cast_fp16 = pow(x = x_25_cast_fp16, y = var_2129_promoted_to_fp16)[name = string("op_2131_cast_fp16")];
            tensor<int32, [1]> var_2133_axes_0 = const()[name = string("op_2133_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2133_keep_dims_0 = const()[name = string("op_2133_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2133_cast_fp16 = reduce_mean(axes = var_2133_axes_0, keep_dims = var_2133_keep_dims_0, x = var_2131_cast_fp16)[name = string("op_2133_cast_fp16")];
            fp16 var_2134_to_fp16 = const()[name = string("op_2134_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2135_cast_fp16 = add(x = var_2133_cast_fp16, y = var_2134_to_fp16)[name = string("op_2135_cast_fp16")];
            fp32 var_2136_epsilon_0 = const()[name = string("op_2136_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2136_cast_fp16 = rsqrt(epsilon = var_2136_epsilon_0, x = var_2135_cast_fp16)[name = string("op_2136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2137_cast_fp16 = mul(x = x_25_cast_fp16, y = var_2136_cast_fp16)[name = string("op_2137_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342766592)))];
            tensor<fp16, [1, 16, 1024]> input_21_cast_fp16 = mul(x = var_2137_cast_fp16, y = model_layers_2_input_layernorm_weight_to_fp16)[name = string("input_21_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_2_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(342768704))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344865920))))[name = string("model_layers_2_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_14_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_2_self_attn_q_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_14_cast_fp16")];
            tensor<int32, [4]> var_2146 = const()[name = string("op_2146"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2147_cast_fp16 = reshape(shape = var_2146, x = linear_14_cast_fp16)[name = string("op_2147_cast_fp16")];
            tensor<int32, [4]> x_27_perm_0 = const()[name = string("x_27_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2155_promoted_to_fp16 = const()[name = string("op_2155_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_27_cast_fp16 = transpose(perm = x_27_perm_0, x = var_2147_cast_fp16)[name = string("transpose_129")];
            tensor<fp16, [1, 16, 16, 128]> var_2157_cast_fp16 = pow(x = x_27_cast_fp16, y = var_2155_promoted_to_fp16)[name = string("op_2157_cast_fp16")];
            tensor<int32, [1]> var_2159_axes_0 = const()[name = string("op_2159_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2159_keep_dims_0 = const()[name = string("op_2159_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2159_cast_fp16 = reduce_mean(axes = var_2159_axes_0, keep_dims = var_2159_keep_dims_0, x = var_2157_cast_fp16)[name = string("op_2159_cast_fp16")];
            fp16 var_2160_to_fp16 = const()[name = string("op_2160_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2161_cast_fp16 = add(x = var_2159_cast_fp16, y = var_2160_to_fp16)[name = string("op_2161_cast_fp16")];
            fp32 var_2162_epsilon_0 = const()[name = string("op_2162_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2162_cast_fp16 = rsqrt(epsilon = var_2162_epsilon_0, x = var_2161_cast_fp16)[name = string("op_2162_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2163_cast_fp16 = mul(x = x_27_cast_fp16, y = var_2162_cast_fp16)[name = string("op_2163_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870080)))];
            tensor<fp16, [1, 16, 16, 128]> x_31_cast_fp16 = mul(x = var_2163_cast_fp16, y = model_layers_2_self_attn_q_norm_weight_to_fp16)[name = string("x_31_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(344870400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345919040))))[name = string("model_layers_2_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_15_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_k_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_15_cast_fp16")];
            tensor<int32, [4]> var_2172 = const()[name = string("op_2172"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2173_cast_fp16 = reshape(shape = var_2172, x = linear_15_cast_fp16)[name = string("op_2173_cast_fp16")];
            tensor<int32, [4]> x_29_perm_0 = const()[name = string("x_29_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2181_promoted_to_fp16 = const()[name = string("op_2181_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_29_cast_fp16 = transpose(perm = x_29_perm_0, x = var_2173_cast_fp16)[name = string("transpose_128")];
            tensor<fp16, [1, 8, 16, 128]> var_2183_cast_fp16 = pow(x = x_29_cast_fp16, y = var_2181_promoted_to_fp16)[name = string("op_2183_cast_fp16")];
            tensor<int32, [1]> var_2185_axes_0 = const()[name = string("op_2185_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2185_keep_dims_0 = const()[name = string("op_2185_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2185_cast_fp16 = reduce_mean(axes = var_2185_axes_0, keep_dims = var_2185_keep_dims_0, x = var_2183_cast_fp16)[name = string("op_2185_cast_fp16")];
            fp16 var_2186_to_fp16 = const()[name = string("op_2186_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2187_cast_fp16 = add(x = var_2185_cast_fp16, y = var_2186_to_fp16)[name = string("op_2187_cast_fp16")];
            fp32 var_2188_epsilon_0 = const()[name = string("op_2188_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2188_cast_fp16 = rsqrt(epsilon = var_2188_epsilon_0, x = var_2187_cast_fp16)[name = string("op_2188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2189_cast_fp16 = mul(x = x_29_cast_fp16, y = var_2188_cast_fp16)[name = string("op_2189_cast_fp16")];
            tensor<fp16, [128]> model_layers_2_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_2_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921152)))];
            tensor<fp16, [1, 8, 16, 128]> x_33_cast_fp16 = mul(x = var_2189_cast_fp16, y = model_layers_2_self_attn_k_norm_weight_to_fp16)[name = string("x_33_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_2_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(345921472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346970112))))[name = string("model_layers_2_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_16_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_v_proj_weight_to_fp16_quantized, x = input_21_cast_fp16)[name = string("linear_16_cast_fp16")];
            tensor<int32, [4]> var_2198 = const()[name = string("op_2198"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2199_cast_fp16 = reshape(shape = var_2198, x = linear_16_cast_fp16)[name = string("op_2199_cast_fp16")];
            tensor<int32, [4]> v_5_perm_0 = const()[name = string("v_5_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_9_begin_0 = const()[name = string("first_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_9_end_0 = const()[name = string("first_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_9_end_mask_0 = const()[name = string("first_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_9_cast_fp16 = slice_by_index(begin = first_9_begin_0, end = first_9_end_0, end_mask = first_9_end_mask_0, x = x_31_cast_fp16)[name = string("first_9_cast_fp16")];
            tensor<int32, [4]> second_9_begin_0 = const()[name = string("second_9_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_9_end_0 = const()[name = string("second_9_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_9_end_mask_0 = const()[name = string("second_9_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_9_cast_fp16 = slice_by_index(begin = second_9_begin_0, end = second_9_end_0, end_mask = second_9_end_mask_0, x = x_31_cast_fp16)[name = string("second_9_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2213_cast_fp16 = mul(x = first_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2213_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2214_cast_fp16 = mul(x = second_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2214_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2216_cast_fp16 = sub(x = var_2213_cast_fp16, y = var_2214_cast_fp16)[name = string("op_2216_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2217_cast_fp16 = mul(x = second_9_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2217_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2218_cast_fp16 = mul(x = first_9_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2218_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2220_cast_fp16 = add(x = var_2217_cast_fp16, y = var_2218_cast_fp16)[name = string("op_2220_cast_fp16")];
            int32 var_2222 = const()[name = string("op_2222"), val = int32(-1)];
            bool q_5_interleave_0 = const()[name = string("q_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_5_cast_fp16 = concat(axis = var_2222, interleave = q_5_interleave_0, values = (var_2216_cast_fp16, var_2220_cast_fp16))[name = string("q_5_cast_fp16")];
            tensor<int32, [4]> first_11_begin_0 = const()[name = string("first_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_11_end_0 = const()[name = string("first_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_11_end_mask_0 = const()[name = string("first_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_11_cast_fp16 = slice_by_index(begin = first_11_begin_0, end = first_11_end_0, end_mask = first_11_end_mask_0, x = x_33_cast_fp16)[name = string("first_11_cast_fp16")];
            tensor<int32, [4]> second_11_begin_0 = const()[name = string("second_11_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_11_end_0 = const()[name = string("second_11_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_11_end_mask_0 = const()[name = string("second_11_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_11_cast_fp16 = slice_by_index(begin = second_11_begin_0, end = second_11_end_0, end_mask = second_11_end_mask_0, x = x_33_cast_fp16)[name = string("second_11_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2234_cast_fp16 = mul(x = first_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2234_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2235_cast_fp16 = mul(x = second_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2235_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2237_cast_fp16 = sub(x = var_2234_cast_fp16, y = var_2235_cast_fp16)[name = string("op_2237_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2238_cast_fp16 = mul(x = second_11_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2238_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2239_cast_fp16 = mul(x = first_11_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2239_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2241_cast_fp16 = add(x = var_2238_cast_fp16, y = var_2239_cast_fp16)[name = string("op_2241_cast_fp16")];
            int32 var_2243 = const()[name = string("op_2243"), val = int32(-1)];
            bool k_5_interleave_0 = const()[name = string("k_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_5_cast_fp16 = concat(axis = var_2243, interleave = k_5_interleave_0, values = (var_2237_cast_fp16, var_2241_cast_fp16))[name = string("k_5_cast_fp16")];
            tensor<int32, [4]> var_2254_begin_0 = const()[name = string("op_2254_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2254_end_0 = const()[name = string("op_2254_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2254_end_mask_0 = const()[name = string("op_2254_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2254_stride_0 = const()[name = string("op_2254_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2254_cast_fp16 = slice_by_index(begin = var_2254_begin_0, end = var_2254_end_0, end_mask = var_2254_end_mask_0, stride = var_2254_stride_0, x = q_5_cast_fp16)[name = string("op_2254_cast_fp16")];
            tensor<int32, [4]> var_2264_begin_0 = const()[name = string("op_2264_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2264_end_0 = const()[name = string("op_2264_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2264_end_mask_0 = const()[name = string("op_2264_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2264_stride_0 = const()[name = string("op_2264_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2264_cast_fp16 = slice_by_index(begin = var_2264_begin_0, end = var_2264_end_0, end_mask = var_2264_end_mask_0, stride = var_2264_stride_0, x = q_5_cast_fp16)[name = string("op_2264_cast_fp16")];
            int32 var_2266 = const()[name = string("op_2266"), val = int32(2)];
            bool stacked_5_interleave_0 = const()[name = string("stacked_5_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_5_cast_fp16 = concat(axis = var_2266, interleave = stacked_5_interleave_0, values = (var_2254_cast_fp16, var_2264_cast_fp16))[name = string("stacked_5_cast_fp16")];
            tensor<int32, [4]> var_2272_begin_0 = const()[name = string("op_2272_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2272_end_0 = const()[name = string("op_2272_end_0"), val = tensor<int32, [4]>([3, 8, 528, 128])];
            tensor<bool, [4]> var_2272_end_mask_0 = const()[name = string("op_2272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2272_cast_fp16 = slice_by_index(begin = var_2272_begin_0, end = var_2272_end_0, end_mask = var_2272_end_mask_0, x = read_state_0)[name = string("op_2272_cast_fp16")];
            tensor<int32, [4]> var_2282_begin_0 = const()[name = string("op_2282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2282_end_0 = const()[name = string("op_2282_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2282_end_mask_0 = const()[name = string("op_2282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2282_cast_fp16 = slice_by_index(begin = var_2282_begin_0, end = var_2282_end_0, end_mask = var_2282_end_mask_0, x = var_2272_cast_fp16)[name = string("op_2282_cast_fp16")];
            bool var_2291_transpose_x_1 = const()[name = string("op_2291_transpose_x_1"), val = bool(false)];
            bool var_2291_transpose_y_1 = const()[name = string("op_2291_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2291_cast_fp16 = matmul(transpose_x = var_2291_transpose_x_1, transpose_y = var_2291_transpose_y_1, x = var_2282_cast_fp16, y = stacked_5_cast_fp16)[name = string("op_2291_cast_fp16")];
            tensor<int32, [4]> var_2294_perm_0 = const()[name = string("op_2294_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2295_to_fp16 = const()[name = string("op_2295_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2294_cast_fp16 = transpose(perm = var_2294_perm_0, x = var_2291_cast_fp16)[name = string("transpose_127")];
            tensor<fp16, [1, 8, 32, 512]> var_2296_cast_fp16 = mul(x = var_2294_cast_fp16, y = var_2295_to_fp16)[name = string("op_2296_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_5_cast_fp16 = add(x = var_2296_cast_fp16, y = cacheMask)[name = string("cache_weights_5_cast_fp16")];
            bool var_2302_transpose_x_1 = const()[name = string("op_2302_transpose_x_1"), val = bool(false)];
            bool var_2302_transpose_y_1 = const()[name = string("op_2302_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2302_cast_fp16 = matmul(transpose_x = var_2302_transpose_x_1, transpose_y = var_2302_transpose_y_1, x = stacked_5_cast_fp16, y = k_5_cast_fp16)[name = string("op_2302_cast_fp16")];
            fp16 var_2303_to_fp16 = const()[name = string("op_2303_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2304_cast_fp16 = mul(x = var_2302_cast_fp16, y = var_2303_to_fp16)[name = string("op_2304_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_5_cast_fp16 = add(x = var_2304_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_5_cast_fp16")];
            int32 var_2308 = const()[name = string("op_2308"), val = int32(-1)];
            bool var_2309_interleave_0 = const()[name = string("op_2309_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2309_cast_fp16 = concat(axis = var_2308, interleave = var_2309_interleave_0, values = (cache_weights_5_cast_fp16, call_weights_5_cast_fp16))[name = string("op_2309_cast_fp16")];
            int32 var_2310 = const()[name = string("op_2310"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_5_cast_fp16 = softmax(axis = var_2310, x = var_2309_cast_fp16)[name = string("shares_5_cast_fp16")];
            tensor<int32, [4]> var_2317_begin_0 = const()[name = string("op_2317_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2317_end_0 = const()[name = string("op_2317_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2317_end_mask_0 = const()[name = string("op_2317_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2317_cast_fp16 = slice_by_index(begin = var_2317_begin_0, end = var_2317_end_0, end_mask = var_2317_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2317_cast_fp16")];
            tensor<int32, [4]> var_2322_begin_0 = const()[name = string("op_2322_begin_0"), val = tensor<int32, [4]>([2, 0, 0, 0])];
            tensor<int32, [4]> var_2322_end_0 = const()[name = string("op_2322_end_0"), val = tensor<int32, [4]>([3, 8, 528, 128])];
            tensor<bool, [4]> var_2322_end_mask_0 = const()[name = string("op_2322_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2322_cast_fp16 = slice_by_index(begin = var_2322_begin_0, end = var_2322_end_0, end_mask = var_2322_end_mask_0, x = read_state_1)[name = string("op_2322_cast_fp16")];
            tensor<int32, [4]> var_2332_begin_0 = const()[name = string("op_2332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2332_end_0 = const()[name = string("op_2332_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2332_end_mask_0 = const()[name = string("op_2332_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2332_cast_fp16 = slice_by_index(begin = var_2332_begin_0, end = var_2332_end_0, end_mask = var_2332_end_mask_0, x = var_2322_cast_fp16)[name = string("op_2332_cast_fp16")];
            bool var_2338_transpose_x_0 = const()[name = string("op_2338_transpose_x_0"), val = bool(false)];
            bool var_2338_transpose_y_0 = const()[name = string("op_2338_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2338_cast_fp16 = matmul(transpose_x = var_2338_transpose_x_0, transpose_y = var_2338_transpose_y_0, x = var_2317_cast_fp16, y = var_2332_cast_fp16)[name = string("op_2338_cast_fp16")];
            tensor<int32, [4]> var_2343_begin_0 = const()[name = string("op_2343_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2343_end_0 = const()[name = string("op_2343_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2343_end_mask_0 = const()[name = string("op_2343_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2343_cast_fp16 = slice_by_index(begin = var_2343_begin_0, end = var_2343_end_0, end_mask = var_2343_end_mask_0, x = shares_5_cast_fp16)[name = string("op_2343_cast_fp16")];
            bool var_2344_transpose_x_0 = const()[name = string("op_2344_transpose_x_0"), val = bool(false)];
            bool var_2344_transpose_y_0 = const()[name = string("op_2344_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_5_cast_fp16 = transpose(perm = v_5_perm_0, x = var_2199_cast_fp16)[name = string("transpose_126")];
            tensor<fp16, [1, 8, 32, 128]> var_2344_cast_fp16 = matmul(transpose_x = var_2344_transpose_x_0, transpose_y = var_2344_transpose_y_0, x = var_2343_cast_fp16, y = v_5_cast_fp16)[name = string("op_2344_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_9_cast_fp16 = add(x = var_2338_cast_fp16, y = var_2344_cast_fp16)[name = string("mixed_9_cast_fp16")];
            tensor<int32, [4]> var_2361_begin_0 = const()[name = string("op_2361_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2361_end_0 = const()[name = string("op_2361_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2361_end_mask_0 = const()[name = string("op_2361_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2361_cast_fp16 = slice_by_index(begin = var_2361_begin_0, end = var_2361_end_0, end_mask = var_2361_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2361_cast_fp16")];
            tensor<int32, [4]> var_2376_begin_0 = const()[name = string("op_2376_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2376_end_0 = const()[name = string("op_2376_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2376_end_mask_0 = const()[name = string("op_2376_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2376_cast_fp16 = slice_by_index(begin = var_2376_begin_0, end = var_2376_end_0, end_mask = var_2376_end_mask_0, x = mixed_9_cast_fp16)[name = string("op_2376_cast_fp16")];
            int32 var_2378 = const()[name = string("op_2378"), val = int32(1)];
            bool mixed_11_interleave_0 = const()[name = string("mixed_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_11_cast_fp16 = concat(axis = var_2378, interleave = mixed_11_interleave_0, values = (var_2361_cast_fp16, var_2376_cast_fp16))[name = string("mixed_11_cast_fp16")];
            tensor<int32, [4]> var_2382_perm_0 = const()[name = string("op_2382_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2386 = const()[name = string("op_2386"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2382_cast_fp16 = transpose(perm = var_2382_perm_0, x = mixed_11_cast_fp16)[name = string("transpose_125")];
            tensor<fp16, [1, 16, 2048]> input_23_cast_fp16 = reshape(shape = var_2386, x = var_2382_cast_fp16)[name = string("input_23_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_2_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(346972224))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349069440))))[name = string("model_layers_2_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_17_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_self_attn_o_proj_weight_to_fp16_quantized, x = input_23_cast_fp16)[name = string("linear_17_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_35_cast_fp16 = add(x = x_25_cast_fp16, y = linear_17_cast_fp16)[name = string("x_35_cast_fp16")];
            fp16 var_2397_promoted_to_fp16 = const()[name = string("op_2397_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2399_cast_fp16 = pow(x = x_35_cast_fp16, y = var_2397_promoted_to_fp16)[name = string("op_2399_cast_fp16")];
            tensor<int32, [1]> var_2401_axes_0 = const()[name = string("op_2401_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2401_keep_dims_0 = const()[name = string("op_2401_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2401_cast_fp16 = reduce_mean(axes = var_2401_axes_0, keep_dims = var_2401_keep_dims_0, x = var_2399_cast_fp16)[name = string("op_2401_cast_fp16")];
            fp16 var_2402_to_fp16 = const()[name = string("op_2402_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2403_cast_fp16 = add(x = var_2401_cast_fp16, y = var_2402_to_fp16)[name = string("op_2403_cast_fp16")];
            fp32 var_2404_epsilon_0 = const()[name = string("op_2404_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2404_cast_fp16 = rsqrt(epsilon = var_2404_epsilon_0, x = var_2403_cast_fp16)[name = string("op_2404_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2405_cast_fp16 = mul(x = x_35_cast_fp16, y = var_2404_cast_fp16)[name = string("op_2405_cast_fp16")];
            tensor<fp16, [1024]> model_layers_2_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_2_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349071552)))];
            tensor<fp16, [1, 16, 1024]> input_25_cast_fp16 = mul(x = var_2405_cast_fp16, y = model_layers_2_post_attention_layernorm_weight_to_fp16)[name = string("input_25_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(349073664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352219456))))[name = string("model_layers_2_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_18_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_gate_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_18_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2410_cast_fp16 = silu(x = linear_18_cast_fp16)[name = string("op_2410_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_2_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(352225664))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355371456))))[name = string("model_layers_2_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_19_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_2_mlp_up_proj_weight_to_fp16_quantized, x = input_25_cast_fp16)[name = string("linear_19_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_29_cast_fp16 = mul(x = var_2410_cast_fp16, y = linear_19_cast_fp16)[name = string("input_29_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_2_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(355377664))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358523456))))[name = string("model_layers_2_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_20_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_2_mlp_down_proj_weight_to_fp16_quantized, x = input_29_cast_fp16)[name = string("linear_20_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_37_cast_fp16 = add(x = x_35_cast_fp16, y = linear_20_cast_fp16)[name = string("x_37_cast_fp16")];
            fp16 var_2424_promoted_to_fp16 = const()[name = string("op_2424_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2426_cast_fp16 = pow(x = x_37_cast_fp16, y = var_2424_promoted_to_fp16)[name = string("op_2426_cast_fp16")];
            tensor<int32, [1]> var_2428_axes_0 = const()[name = string("op_2428_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2428_keep_dims_0 = const()[name = string("op_2428_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2428_cast_fp16 = reduce_mean(axes = var_2428_axes_0, keep_dims = var_2428_keep_dims_0, x = var_2426_cast_fp16)[name = string("op_2428_cast_fp16")];
            fp16 var_2429_to_fp16 = const()[name = string("op_2429_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2430_cast_fp16 = add(x = var_2428_cast_fp16, y = var_2429_to_fp16)[name = string("op_2430_cast_fp16")];
            fp32 var_2431_epsilon_0 = const()[name = string("op_2431_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2431_cast_fp16 = rsqrt(epsilon = var_2431_epsilon_0, x = var_2430_cast_fp16)[name = string("op_2431_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2432_cast_fp16 = mul(x = x_37_cast_fp16, y = var_2431_cast_fp16)[name = string("op_2432_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358525568)))];
            tensor<fp16, [1, 16, 1024]> input_31_cast_fp16 = mul(x = var_2432_cast_fp16, y = model_layers_3_input_layernorm_weight_to_fp16)[name = string("input_31_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_3_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(358527680))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360624896))))[name = string("model_layers_3_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_21_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_3_self_attn_q_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_21_cast_fp16")];
            tensor<int32, [4]> var_2441 = const()[name = string("op_2441"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2442_cast_fp16 = reshape(shape = var_2441, x = linear_21_cast_fp16)[name = string("op_2442_cast_fp16")];
            tensor<int32, [4]> x_39_perm_0 = const()[name = string("x_39_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2450_promoted_to_fp16 = const()[name = string("op_2450_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_39_cast_fp16 = transpose(perm = x_39_perm_0, x = var_2442_cast_fp16)[name = string("transpose_124")];
            tensor<fp16, [1, 16, 16, 128]> var_2452_cast_fp16 = pow(x = x_39_cast_fp16, y = var_2450_promoted_to_fp16)[name = string("op_2452_cast_fp16")];
            tensor<int32, [1]> var_2454_axes_0 = const()[name = string("op_2454_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2454_keep_dims_0 = const()[name = string("op_2454_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2454_cast_fp16 = reduce_mean(axes = var_2454_axes_0, keep_dims = var_2454_keep_dims_0, x = var_2452_cast_fp16)[name = string("op_2454_cast_fp16")];
            fp16 var_2455_to_fp16 = const()[name = string("op_2455_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2456_cast_fp16 = add(x = var_2454_cast_fp16, y = var_2455_to_fp16)[name = string("op_2456_cast_fp16")];
            fp32 var_2457_epsilon_0 = const()[name = string("op_2457_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2457_cast_fp16 = rsqrt(epsilon = var_2457_epsilon_0, x = var_2456_cast_fp16)[name = string("op_2457_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2458_cast_fp16 = mul(x = x_39_cast_fp16, y = var_2457_cast_fp16)[name = string("op_2458_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629056)))];
            tensor<fp16, [1, 16, 16, 128]> x_43_cast_fp16 = mul(x = var_2458_cast_fp16, y = model_layers_3_self_attn_q_norm_weight_to_fp16)[name = string("x_43_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(360629376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361678016))))[name = string("model_layers_3_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_22_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_k_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_22_cast_fp16")];
            tensor<int32, [4]> var_2467 = const()[name = string("op_2467"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2468_cast_fp16 = reshape(shape = var_2467, x = linear_22_cast_fp16)[name = string("op_2468_cast_fp16")];
            tensor<int32, [4]> x_41_perm_0 = const()[name = string("x_41_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2476_promoted_to_fp16 = const()[name = string("op_2476_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_41_cast_fp16 = transpose(perm = x_41_perm_0, x = var_2468_cast_fp16)[name = string("transpose_123")];
            tensor<fp16, [1, 8, 16, 128]> var_2478_cast_fp16 = pow(x = x_41_cast_fp16, y = var_2476_promoted_to_fp16)[name = string("op_2478_cast_fp16")];
            tensor<int32, [1]> var_2480_axes_0 = const()[name = string("op_2480_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2480_keep_dims_0 = const()[name = string("op_2480_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2480_cast_fp16 = reduce_mean(axes = var_2480_axes_0, keep_dims = var_2480_keep_dims_0, x = var_2478_cast_fp16)[name = string("op_2480_cast_fp16")];
            fp16 var_2481_to_fp16 = const()[name = string("op_2481_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2482_cast_fp16 = add(x = var_2480_cast_fp16, y = var_2481_to_fp16)[name = string("op_2482_cast_fp16")];
            fp32 var_2483_epsilon_0 = const()[name = string("op_2483_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2483_cast_fp16 = rsqrt(epsilon = var_2483_epsilon_0, x = var_2482_cast_fp16)[name = string("op_2483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2484_cast_fp16 = mul(x = x_41_cast_fp16, y = var_2483_cast_fp16)[name = string("op_2484_cast_fp16")];
            tensor<fp16, [128]> model_layers_3_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_3_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680128)))];
            tensor<fp16, [1, 8, 16, 128]> x_45_cast_fp16 = mul(x = var_2484_cast_fp16, y = model_layers_3_self_attn_k_norm_weight_to_fp16)[name = string("x_45_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_3_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(361680448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362729088))))[name = string("model_layers_3_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_23_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_v_proj_weight_to_fp16_quantized, x = input_31_cast_fp16)[name = string("linear_23_cast_fp16")];
            tensor<int32, [4]> var_2493 = const()[name = string("op_2493"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2494_cast_fp16 = reshape(shape = var_2493, x = linear_23_cast_fp16)[name = string("op_2494_cast_fp16")];
            tensor<int32, [4]> v_7_perm_0 = const()[name = string("v_7_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_13_begin_0 = const()[name = string("first_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_13_end_0 = const()[name = string("first_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_13_end_mask_0 = const()[name = string("first_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_13_cast_fp16 = slice_by_index(begin = first_13_begin_0, end = first_13_end_0, end_mask = first_13_end_mask_0, x = x_43_cast_fp16)[name = string("first_13_cast_fp16")];
            tensor<int32, [4]> second_13_begin_0 = const()[name = string("second_13_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_13_end_0 = const()[name = string("second_13_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_13_end_mask_0 = const()[name = string("second_13_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_13_cast_fp16 = slice_by_index(begin = second_13_begin_0, end = second_13_end_0, end_mask = second_13_end_mask_0, x = x_43_cast_fp16)[name = string("second_13_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2508_cast_fp16 = mul(x = first_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2508_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2509_cast_fp16 = mul(x = second_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2509_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2511_cast_fp16 = sub(x = var_2508_cast_fp16, y = var_2509_cast_fp16)[name = string("op_2511_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2512_cast_fp16 = mul(x = second_13_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2512_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2513_cast_fp16 = mul(x = first_13_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2513_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2515_cast_fp16 = add(x = var_2512_cast_fp16, y = var_2513_cast_fp16)[name = string("op_2515_cast_fp16")];
            int32 var_2517 = const()[name = string("op_2517"), val = int32(-1)];
            bool q_7_interleave_0 = const()[name = string("q_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_7_cast_fp16 = concat(axis = var_2517, interleave = q_7_interleave_0, values = (var_2511_cast_fp16, var_2515_cast_fp16))[name = string("q_7_cast_fp16")];
            tensor<int32, [4]> first_15_begin_0 = const()[name = string("first_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_15_end_0 = const()[name = string("first_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_15_end_mask_0 = const()[name = string("first_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_15_cast_fp16 = slice_by_index(begin = first_15_begin_0, end = first_15_end_0, end_mask = first_15_end_mask_0, x = x_45_cast_fp16)[name = string("first_15_cast_fp16")];
            tensor<int32, [4]> second_15_begin_0 = const()[name = string("second_15_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_15_end_0 = const()[name = string("second_15_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_15_end_mask_0 = const()[name = string("second_15_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_15_cast_fp16 = slice_by_index(begin = second_15_begin_0, end = second_15_end_0, end_mask = second_15_end_mask_0, x = x_45_cast_fp16)[name = string("second_15_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2529_cast_fp16 = mul(x = first_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2529_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2530_cast_fp16 = mul(x = second_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2530_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2532_cast_fp16 = sub(x = var_2529_cast_fp16, y = var_2530_cast_fp16)[name = string("op_2532_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2533_cast_fp16 = mul(x = second_15_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2533_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2534_cast_fp16 = mul(x = first_15_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2534_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2536_cast_fp16 = add(x = var_2533_cast_fp16, y = var_2534_cast_fp16)[name = string("op_2536_cast_fp16")];
            int32 var_2538 = const()[name = string("op_2538"), val = int32(-1)];
            bool k_7_interleave_0 = const()[name = string("k_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_7_cast_fp16 = concat(axis = var_2538, interleave = k_7_interleave_0, values = (var_2532_cast_fp16, var_2536_cast_fp16))[name = string("k_7_cast_fp16")];
            tensor<int32, [4]> var_2549_begin_0 = const()[name = string("op_2549_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2549_end_0 = const()[name = string("op_2549_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2549_end_mask_0 = const()[name = string("op_2549_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2549_stride_0 = const()[name = string("op_2549_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2549_cast_fp16 = slice_by_index(begin = var_2549_begin_0, end = var_2549_end_0, end_mask = var_2549_end_mask_0, stride = var_2549_stride_0, x = q_7_cast_fp16)[name = string("op_2549_cast_fp16")];
            tensor<int32, [4]> var_2559_begin_0 = const()[name = string("op_2559_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2559_end_0 = const()[name = string("op_2559_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2559_end_mask_0 = const()[name = string("op_2559_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2559_stride_0 = const()[name = string("op_2559_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2559_cast_fp16 = slice_by_index(begin = var_2559_begin_0, end = var_2559_end_0, end_mask = var_2559_end_mask_0, stride = var_2559_stride_0, x = q_7_cast_fp16)[name = string("op_2559_cast_fp16")];
            int32 var_2561 = const()[name = string("op_2561"), val = int32(2)];
            bool stacked_7_interleave_0 = const()[name = string("stacked_7_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_7_cast_fp16 = concat(axis = var_2561, interleave = stacked_7_interleave_0, values = (var_2549_cast_fp16, var_2559_cast_fp16))[name = string("stacked_7_cast_fp16")];
            tensor<int32, [4]> var_2567_begin_0 = const()[name = string("op_2567_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2567_end_0 = const()[name = string("op_2567_end_0"), val = tensor<int32, [4]>([4, 8, 528, 128])];
            tensor<bool, [4]> var_2567_end_mask_0 = const()[name = string("op_2567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2567_cast_fp16 = slice_by_index(begin = var_2567_begin_0, end = var_2567_end_0, end_mask = var_2567_end_mask_0, x = read_state_0)[name = string("op_2567_cast_fp16")];
            tensor<int32, [4]> var_2577_begin_0 = const()[name = string("op_2577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2577_end_0 = const()[name = string("op_2577_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2577_end_mask_0 = const()[name = string("op_2577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2577_cast_fp16 = slice_by_index(begin = var_2577_begin_0, end = var_2577_end_0, end_mask = var_2577_end_mask_0, x = var_2567_cast_fp16)[name = string("op_2577_cast_fp16")];
            bool var_2586_transpose_x_1 = const()[name = string("op_2586_transpose_x_1"), val = bool(false)];
            bool var_2586_transpose_y_1 = const()[name = string("op_2586_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2586_cast_fp16 = matmul(transpose_x = var_2586_transpose_x_1, transpose_y = var_2586_transpose_y_1, x = var_2577_cast_fp16, y = stacked_7_cast_fp16)[name = string("op_2586_cast_fp16")];
            tensor<int32, [4]> var_2589_perm_0 = const()[name = string("op_2589_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2590_to_fp16 = const()[name = string("op_2590_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2589_cast_fp16 = transpose(perm = var_2589_perm_0, x = var_2586_cast_fp16)[name = string("transpose_122")];
            tensor<fp16, [1, 8, 32, 512]> var_2591_cast_fp16 = mul(x = var_2589_cast_fp16, y = var_2590_to_fp16)[name = string("op_2591_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_7_cast_fp16 = add(x = var_2591_cast_fp16, y = cacheMask)[name = string("cache_weights_7_cast_fp16")];
            bool var_2597_transpose_x_1 = const()[name = string("op_2597_transpose_x_1"), val = bool(false)];
            bool var_2597_transpose_y_1 = const()[name = string("op_2597_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2597_cast_fp16 = matmul(transpose_x = var_2597_transpose_x_1, transpose_y = var_2597_transpose_y_1, x = stacked_7_cast_fp16, y = k_7_cast_fp16)[name = string("op_2597_cast_fp16")];
            fp16 var_2598_to_fp16 = const()[name = string("op_2598_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2599_cast_fp16 = mul(x = var_2597_cast_fp16, y = var_2598_to_fp16)[name = string("op_2599_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_7_cast_fp16 = add(x = var_2599_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_7_cast_fp16")];
            int32 var_2603 = const()[name = string("op_2603"), val = int32(-1)];
            bool var_2604_interleave_0 = const()[name = string("op_2604_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2604_cast_fp16 = concat(axis = var_2603, interleave = var_2604_interleave_0, values = (cache_weights_7_cast_fp16, call_weights_7_cast_fp16))[name = string("op_2604_cast_fp16")];
            int32 var_2605 = const()[name = string("op_2605"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_7_cast_fp16 = softmax(axis = var_2605, x = var_2604_cast_fp16)[name = string("shares_7_cast_fp16")];
            tensor<int32, [4]> var_2612_begin_0 = const()[name = string("op_2612_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2612_end_0 = const()[name = string("op_2612_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2612_end_mask_0 = const()[name = string("op_2612_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2612_cast_fp16 = slice_by_index(begin = var_2612_begin_0, end = var_2612_end_0, end_mask = var_2612_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2612_cast_fp16")];
            tensor<int32, [4]> var_2617_begin_0 = const()[name = string("op_2617_begin_0"), val = tensor<int32, [4]>([3, 0, 0, 0])];
            tensor<int32, [4]> var_2617_end_0 = const()[name = string("op_2617_end_0"), val = tensor<int32, [4]>([4, 8, 528, 128])];
            tensor<bool, [4]> var_2617_end_mask_0 = const()[name = string("op_2617_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2617_cast_fp16 = slice_by_index(begin = var_2617_begin_0, end = var_2617_end_0, end_mask = var_2617_end_mask_0, x = read_state_1)[name = string("op_2617_cast_fp16")];
            tensor<int32, [4]> var_2627_begin_0 = const()[name = string("op_2627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2627_end_0 = const()[name = string("op_2627_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2627_end_mask_0 = const()[name = string("op_2627_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2627_cast_fp16 = slice_by_index(begin = var_2627_begin_0, end = var_2627_end_0, end_mask = var_2627_end_mask_0, x = var_2617_cast_fp16)[name = string("op_2627_cast_fp16")];
            bool var_2633_transpose_x_0 = const()[name = string("op_2633_transpose_x_0"), val = bool(false)];
            bool var_2633_transpose_y_0 = const()[name = string("op_2633_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2633_cast_fp16 = matmul(transpose_x = var_2633_transpose_x_0, transpose_y = var_2633_transpose_y_0, x = var_2612_cast_fp16, y = var_2627_cast_fp16)[name = string("op_2633_cast_fp16")];
            tensor<int32, [4]> var_2638_begin_0 = const()[name = string("op_2638_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2638_end_0 = const()[name = string("op_2638_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2638_end_mask_0 = const()[name = string("op_2638_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2638_cast_fp16 = slice_by_index(begin = var_2638_begin_0, end = var_2638_end_0, end_mask = var_2638_end_mask_0, x = shares_7_cast_fp16)[name = string("op_2638_cast_fp16")];
            bool var_2639_transpose_x_0 = const()[name = string("op_2639_transpose_x_0"), val = bool(false)];
            bool var_2639_transpose_y_0 = const()[name = string("op_2639_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_7_cast_fp16 = transpose(perm = v_7_perm_0, x = var_2494_cast_fp16)[name = string("transpose_121")];
            tensor<fp16, [1, 8, 32, 128]> var_2639_cast_fp16 = matmul(transpose_x = var_2639_transpose_x_0, transpose_y = var_2639_transpose_y_0, x = var_2638_cast_fp16, y = v_7_cast_fp16)[name = string("op_2639_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_13_cast_fp16 = add(x = var_2633_cast_fp16, y = var_2639_cast_fp16)[name = string("mixed_13_cast_fp16")];
            tensor<int32, [4]> var_2656_begin_0 = const()[name = string("op_2656_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2656_end_0 = const()[name = string("op_2656_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2656_end_mask_0 = const()[name = string("op_2656_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2656_cast_fp16 = slice_by_index(begin = var_2656_begin_0, end = var_2656_end_0, end_mask = var_2656_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2656_cast_fp16")];
            tensor<int32, [4]> var_2671_begin_0 = const()[name = string("op_2671_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2671_end_0 = const()[name = string("op_2671_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2671_end_mask_0 = const()[name = string("op_2671_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2671_cast_fp16 = slice_by_index(begin = var_2671_begin_0, end = var_2671_end_0, end_mask = var_2671_end_mask_0, x = mixed_13_cast_fp16)[name = string("op_2671_cast_fp16")];
            int32 var_2673 = const()[name = string("op_2673"), val = int32(1)];
            bool mixed_15_interleave_0 = const()[name = string("mixed_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_15_cast_fp16 = concat(axis = var_2673, interleave = mixed_15_interleave_0, values = (var_2656_cast_fp16, var_2671_cast_fp16))[name = string("mixed_15_cast_fp16")];
            tensor<int32, [4]> var_2677_perm_0 = const()[name = string("op_2677_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2681 = const()[name = string("op_2681"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2677_cast_fp16 = transpose(perm = var_2677_perm_0, x = mixed_15_cast_fp16)[name = string("transpose_120")];
            tensor<fp16, [1, 16, 2048]> input_33_cast_fp16 = reshape(shape = var_2681, x = var_2677_cast_fp16)[name = string("input_33_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_3_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(362731200))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364828416))))[name = string("model_layers_3_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_24_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_self_attn_o_proj_weight_to_fp16_quantized, x = input_33_cast_fp16)[name = string("linear_24_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_47_cast_fp16 = add(x = x_37_cast_fp16, y = linear_24_cast_fp16)[name = string("x_47_cast_fp16")];
            fp16 var_2692_promoted_to_fp16 = const()[name = string("op_2692_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2694_cast_fp16 = pow(x = x_47_cast_fp16, y = var_2692_promoted_to_fp16)[name = string("op_2694_cast_fp16")];
            tensor<int32, [1]> var_2696_axes_0 = const()[name = string("op_2696_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2696_keep_dims_0 = const()[name = string("op_2696_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2696_cast_fp16 = reduce_mean(axes = var_2696_axes_0, keep_dims = var_2696_keep_dims_0, x = var_2694_cast_fp16)[name = string("op_2696_cast_fp16")];
            fp16 var_2697_to_fp16 = const()[name = string("op_2697_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2698_cast_fp16 = add(x = var_2696_cast_fp16, y = var_2697_to_fp16)[name = string("op_2698_cast_fp16")];
            fp32 var_2699_epsilon_0 = const()[name = string("op_2699_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2699_cast_fp16 = rsqrt(epsilon = var_2699_epsilon_0, x = var_2698_cast_fp16)[name = string("op_2699_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2700_cast_fp16 = mul(x = x_47_cast_fp16, y = var_2699_cast_fp16)[name = string("op_2700_cast_fp16")];
            tensor<fp16, [1024]> model_layers_3_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_3_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364830528)))];
            tensor<fp16, [1, 16, 1024]> input_35_cast_fp16 = mul(x = var_2700_cast_fp16, y = model_layers_3_post_attention_layernorm_weight_to_fp16)[name = string("input_35_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(364832640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367978432))))[name = string("model_layers_3_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_25_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_gate_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_25_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_2705_cast_fp16 = silu(x = linear_25_cast_fp16)[name = string("op_2705_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_3_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(367984640))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371130432))))[name = string("model_layers_3_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_26_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_3_mlp_up_proj_weight_to_fp16_quantized, x = input_35_cast_fp16)[name = string("linear_26_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_39_cast_fp16 = mul(x = var_2705_cast_fp16, y = linear_26_cast_fp16)[name = string("input_39_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_3_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(371136640))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374282432))))[name = string("model_layers_3_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_27_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_3_mlp_down_proj_weight_to_fp16_quantized, x = input_39_cast_fp16)[name = string("linear_27_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_49_cast_fp16 = add(x = x_47_cast_fp16, y = linear_27_cast_fp16)[name = string("x_49_cast_fp16")];
            fp16 var_2719_promoted_to_fp16 = const()[name = string("op_2719_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2721_cast_fp16 = pow(x = x_49_cast_fp16, y = var_2719_promoted_to_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]>([-1])];
            bool var_2723_keep_dims_0 = const()[name = string("op_2723_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2723_cast_fp16 = reduce_mean(axes = var_2723_axes_0, keep_dims = var_2723_keep_dims_0, x = var_2721_cast_fp16)[name = string("op_2723_cast_fp16")];
            fp16 var_2724_to_fp16 = const()[name = string("op_2724_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2725_cast_fp16 = add(x = var_2723_cast_fp16, y = var_2724_to_fp16)[name = string("op_2725_cast_fp16")];
            fp32 var_2726_epsilon_0 = const()[name = string("op_2726_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2726_cast_fp16 = rsqrt(epsilon = var_2726_epsilon_0, x = var_2725_cast_fp16)[name = string("op_2726_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2727_cast_fp16 = mul(x = x_49_cast_fp16, y = var_2726_cast_fp16)[name = string("op_2727_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374284544)))];
            tensor<fp16, [1, 16, 1024]> input_41_cast_fp16 = mul(x = var_2727_cast_fp16, y = model_layers_4_input_layernorm_weight_to_fp16)[name = string("input_41_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_4_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(374286656))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376383872))))[name = string("model_layers_4_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_28_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_4_self_attn_q_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_28_cast_fp16")];
            tensor<int32, [4]> var_2736 = const()[name = string("op_2736"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_2737_cast_fp16 = reshape(shape = var_2736, x = linear_28_cast_fp16)[name = string("op_2737_cast_fp16")];
            tensor<int32, [4]> x_51_perm_0 = const()[name = string("x_51_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2745_promoted_to_fp16 = const()[name = string("op_2745_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_51_cast_fp16 = transpose(perm = x_51_perm_0, x = var_2737_cast_fp16)[name = string("transpose_119")];
            tensor<fp16, [1, 16, 16, 128]> var_2747_cast_fp16 = pow(x = x_51_cast_fp16, y = var_2745_promoted_to_fp16)[name = string("op_2747_cast_fp16")];
            tensor<int32, [1]> var_2749_axes_0 = const()[name = string("op_2749_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2749_keep_dims_0 = const()[name = string("op_2749_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_2749_cast_fp16 = reduce_mean(axes = var_2749_axes_0, keep_dims = var_2749_keep_dims_0, x = var_2747_cast_fp16)[name = string("op_2749_cast_fp16")];
            fp16 var_2750_to_fp16 = const()[name = string("op_2750_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_2751_cast_fp16 = add(x = var_2749_cast_fp16, y = var_2750_to_fp16)[name = string("op_2751_cast_fp16")];
            fp32 var_2752_epsilon_0 = const()[name = string("op_2752_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_2752_cast_fp16 = rsqrt(epsilon = var_2752_epsilon_0, x = var_2751_cast_fp16)[name = string("op_2752_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_2753_cast_fp16 = mul(x = x_51_cast_fp16, y = var_2752_cast_fp16)[name = string("op_2753_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388032)))];
            tensor<fp16, [1, 16, 16, 128]> x_55_cast_fp16 = mul(x = var_2753_cast_fp16, y = model_layers_4_self_attn_q_norm_weight_to_fp16)[name = string("x_55_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(376388352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377436992))))[name = string("model_layers_4_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_29_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_k_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_29_cast_fp16")];
            tensor<int32, [4]> var_2762 = const()[name = string("op_2762"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2763_cast_fp16 = reshape(shape = var_2762, x = linear_29_cast_fp16)[name = string("op_2763_cast_fp16")];
            tensor<int32, [4]> x_53_perm_0 = const()[name = string("x_53_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_2771_promoted_to_fp16 = const()[name = string("op_2771_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_53_cast_fp16 = transpose(perm = x_53_perm_0, x = var_2763_cast_fp16)[name = string("transpose_118")];
            tensor<fp16, [1, 8, 16, 128]> var_2773_cast_fp16 = pow(x = x_53_cast_fp16, y = var_2771_promoted_to_fp16)[name = string("op_2773_cast_fp16")];
            tensor<int32, [1]> var_2775_axes_0 = const()[name = string("op_2775_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2775_keep_dims_0 = const()[name = string("op_2775_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_2775_cast_fp16 = reduce_mean(axes = var_2775_axes_0, keep_dims = var_2775_keep_dims_0, x = var_2773_cast_fp16)[name = string("op_2775_cast_fp16")];
            fp16 var_2776_to_fp16 = const()[name = string("op_2776_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_2777_cast_fp16 = add(x = var_2775_cast_fp16, y = var_2776_to_fp16)[name = string("op_2777_cast_fp16")];
            fp32 var_2778_epsilon_0 = const()[name = string("op_2778_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_2778_cast_fp16 = rsqrt(epsilon = var_2778_epsilon_0, x = var_2777_cast_fp16)[name = string("op_2778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_2779_cast_fp16 = mul(x = x_53_cast_fp16, y = var_2778_cast_fp16)[name = string("op_2779_cast_fp16")];
            tensor<fp16, [128]> model_layers_4_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_4_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439104)))];
            tensor<fp16, [1, 8, 16, 128]> x_57_cast_fp16 = mul(x = var_2779_cast_fp16, y = model_layers_4_self_attn_k_norm_weight_to_fp16)[name = string("x_57_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_4_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(377439424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378488064))))[name = string("model_layers_4_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_30_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_v_proj_weight_to_fp16_quantized, x = input_41_cast_fp16)[name = string("linear_30_cast_fp16")];
            tensor<int32, [4]> var_2788 = const()[name = string("op_2788"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_2789_cast_fp16 = reshape(shape = var_2788, x = linear_30_cast_fp16)[name = string("op_2789_cast_fp16")];
            tensor<int32, [4]> v_9_perm_0 = const()[name = string("v_9_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_17_begin_0 = const()[name = string("first_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_17_end_0 = const()[name = string("first_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_17_end_mask_0 = const()[name = string("first_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_17_cast_fp16 = slice_by_index(begin = first_17_begin_0, end = first_17_end_0, end_mask = first_17_end_mask_0, x = x_55_cast_fp16)[name = string("first_17_cast_fp16")];
            tensor<int32, [4]> second_17_begin_0 = const()[name = string("second_17_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_17_end_0 = const()[name = string("second_17_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_17_end_mask_0 = const()[name = string("second_17_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_17_cast_fp16 = slice_by_index(begin = second_17_begin_0, end = second_17_end_0, end_mask = second_17_end_mask_0, x = x_55_cast_fp16)[name = string("second_17_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2803_cast_fp16 = mul(x = first_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2803_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2804_cast_fp16 = mul(x = second_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2804_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2806_cast_fp16 = sub(x = var_2803_cast_fp16, y = var_2804_cast_fp16)[name = string("op_2806_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2807_cast_fp16 = mul(x = second_17_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2807_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2808_cast_fp16 = mul(x = first_17_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2808_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_2810_cast_fp16 = add(x = var_2807_cast_fp16, y = var_2808_cast_fp16)[name = string("op_2810_cast_fp16")];
            int32 var_2812 = const()[name = string("op_2812"), val = int32(-1)];
            bool q_9_interleave_0 = const()[name = string("q_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_9_cast_fp16 = concat(axis = var_2812, interleave = q_9_interleave_0, values = (var_2806_cast_fp16, var_2810_cast_fp16))[name = string("q_9_cast_fp16")];
            tensor<int32, [4]> first_19_begin_0 = const()[name = string("first_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_19_end_0 = const()[name = string("first_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_19_end_mask_0 = const()[name = string("first_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_19_cast_fp16 = slice_by_index(begin = first_19_begin_0, end = first_19_end_0, end_mask = first_19_end_mask_0, x = x_57_cast_fp16)[name = string("first_19_cast_fp16")];
            tensor<int32, [4]> second_19_begin_0 = const()[name = string("second_19_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_19_end_0 = const()[name = string("second_19_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_19_end_mask_0 = const()[name = string("second_19_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_19_cast_fp16 = slice_by_index(begin = second_19_begin_0, end = second_19_end_0, end_mask = second_19_end_mask_0, x = x_57_cast_fp16)[name = string("second_19_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2824_cast_fp16 = mul(x = first_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2824_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2825_cast_fp16 = mul(x = second_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2825_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2827_cast_fp16 = sub(x = var_2824_cast_fp16, y = var_2825_cast_fp16)[name = string("op_2827_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2828_cast_fp16 = mul(x = second_19_cast_fp16, y = cos_1_cast_fp16)[name = string("op_2828_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2829_cast_fp16 = mul(x = first_19_cast_fp16, y = sin_1_cast_fp16)[name = string("op_2829_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_2831_cast_fp16 = add(x = var_2828_cast_fp16, y = var_2829_cast_fp16)[name = string("op_2831_cast_fp16")];
            int32 var_2833 = const()[name = string("op_2833"), val = int32(-1)];
            bool k_9_interleave_0 = const()[name = string("k_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_9_cast_fp16 = concat(axis = var_2833, interleave = k_9_interleave_0, values = (var_2827_cast_fp16, var_2831_cast_fp16))[name = string("k_9_cast_fp16")];
            tensor<int32, [4]> var_2844_begin_0 = const()[name = string("op_2844_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2844_end_0 = const()[name = string("op_2844_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2844_end_mask_0 = const()[name = string("op_2844_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2844_stride_0 = const()[name = string("op_2844_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2844_cast_fp16 = slice_by_index(begin = var_2844_begin_0, end = var_2844_end_0, end_mask = var_2844_end_mask_0, stride = var_2844_stride_0, x = q_9_cast_fp16)[name = string("op_2844_cast_fp16")];
            tensor<int32, [4]> var_2854_begin_0 = const()[name = string("op_2854_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_2854_end_0 = const()[name = string("op_2854_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_2854_end_mask_0 = const()[name = string("op_2854_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_2854_stride_0 = const()[name = string("op_2854_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_2854_cast_fp16 = slice_by_index(begin = var_2854_begin_0, end = var_2854_end_0, end_mask = var_2854_end_mask_0, stride = var_2854_stride_0, x = q_9_cast_fp16)[name = string("op_2854_cast_fp16")];
            int32 var_2856 = const()[name = string("op_2856"), val = int32(2)];
            bool stacked_9_interleave_0 = const()[name = string("stacked_9_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_9_cast_fp16 = concat(axis = var_2856, interleave = stacked_9_interleave_0, values = (var_2844_cast_fp16, var_2854_cast_fp16))[name = string("stacked_9_cast_fp16")];
            tensor<int32, [4]> var_2862_begin_0 = const()[name = string("op_2862_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2862_end_0 = const()[name = string("op_2862_end_0"), val = tensor<int32, [4]>([5, 8, 528, 128])];
            tensor<bool, [4]> var_2862_end_mask_0 = const()[name = string("op_2862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2862_cast_fp16 = slice_by_index(begin = var_2862_begin_0, end = var_2862_end_0, end_mask = var_2862_end_mask_0, x = read_state_0)[name = string("op_2862_cast_fp16")];
            tensor<int32, [4]> var_2872_begin_0 = const()[name = string("op_2872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2872_end_0 = const()[name = string("op_2872_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2872_end_mask_0 = const()[name = string("op_2872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2872_cast_fp16 = slice_by_index(begin = var_2872_begin_0, end = var_2872_end_0, end_mask = var_2872_end_mask_0, x = var_2862_cast_fp16)[name = string("op_2872_cast_fp16")];
            bool var_2881_transpose_x_1 = const()[name = string("op_2881_transpose_x_1"), val = bool(false)];
            bool var_2881_transpose_y_1 = const()[name = string("op_2881_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_2881_cast_fp16 = matmul(transpose_x = var_2881_transpose_x_1, transpose_y = var_2881_transpose_y_1, x = var_2872_cast_fp16, y = stacked_9_cast_fp16)[name = string("op_2881_cast_fp16")];
            tensor<int32, [4]> var_2884_perm_0 = const()[name = string("op_2884_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_2885_to_fp16 = const()[name = string("op_2885_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_2884_cast_fp16 = transpose(perm = var_2884_perm_0, x = var_2881_cast_fp16)[name = string("transpose_117")];
            tensor<fp16, [1, 8, 32, 512]> var_2886_cast_fp16 = mul(x = var_2884_cast_fp16, y = var_2885_to_fp16)[name = string("op_2886_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_9_cast_fp16 = add(x = var_2886_cast_fp16, y = cacheMask)[name = string("cache_weights_9_cast_fp16")];
            bool var_2892_transpose_x_1 = const()[name = string("op_2892_transpose_x_1"), val = bool(false)];
            bool var_2892_transpose_y_1 = const()[name = string("op_2892_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_2892_cast_fp16 = matmul(transpose_x = var_2892_transpose_x_1, transpose_y = var_2892_transpose_y_1, x = stacked_9_cast_fp16, y = k_9_cast_fp16)[name = string("op_2892_cast_fp16")];
            fp16 var_2893_to_fp16 = const()[name = string("op_2893_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_2894_cast_fp16 = mul(x = var_2892_cast_fp16, y = var_2893_to_fp16)[name = string("op_2894_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_9_cast_fp16 = add(x = var_2894_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_9_cast_fp16")];
            int32 var_2898 = const()[name = string("op_2898"), val = int32(-1)];
            bool var_2899_interleave_0 = const()[name = string("op_2899_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_2899_cast_fp16 = concat(axis = var_2898, interleave = var_2899_interleave_0, values = (cache_weights_9_cast_fp16, call_weights_9_cast_fp16))[name = string("op_2899_cast_fp16")];
            int32 var_2900 = const()[name = string("op_2900"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_9_cast_fp16 = softmax(axis = var_2900, x = var_2899_cast_fp16)[name = string("shares_9_cast_fp16")];
            tensor<int32, [4]> var_2907_begin_0 = const()[name = string("op_2907_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2907_end_0 = const()[name = string("op_2907_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_2907_end_mask_0 = const()[name = string("op_2907_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_2907_cast_fp16 = slice_by_index(begin = var_2907_begin_0, end = var_2907_end_0, end_mask = var_2907_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2907_cast_fp16")];
            tensor<int32, [4]> var_2912_begin_0 = const()[name = string("op_2912_begin_0"), val = tensor<int32, [4]>([4, 0, 0, 0])];
            tensor<int32, [4]> var_2912_end_0 = const()[name = string("op_2912_end_0"), val = tensor<int32, [4]>([5, 8, 528, 128])];
            tensor<bool, [4]> var_2912_end_mask_0 = const()[name = string("op_2912_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_2912_cast_fp16 = slice_by_index(begin = var_2912_begin_0, end = var_2912_end_0, end_mask = var_2912_end_mask_0, x = read_state_1)[name = string("op_2912_cast_fp16")];
            tensor<int32, [4]> var_2922_begin_0 = const()[name = string("op_2922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2922_end_0 = const()[name = string("op_2922_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_2922_end_mask_0 = const()[name = string("op_2922_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_2922_cast_fp16 = slice_by_index(begin = var_2922_begin_0, end = var_2922_end_0, end_mask = var_2922_end_mask_0, x = var_2912_cast_fp16)[name = string("op_2922_cast_fp16")];
            bool var_2928_transpose_x_0 = const()[name = string("op_2928_transpose_x_0"), val = bool(false)];
            bool var_2928_transpose_y_0 = const()[name = string("op_2928_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_2928_cast_fp16 = matmul(transpose_x = var_2928_transpose_x_0, transpose_y = var_2928_transpose_y_0, x = var_2907_cast_fp16, y = var_2922_cast_fp16)[name = string("op_2928_cast_fp16")];
            tensor<int32, [4]> var_2933_begin_0 = const()[name = string("op_2933_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_2933_end_0 = const()[name = string("op_2933_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_2933_end_mask_0 = const()[name = string("op_2933_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_2933_cast_fp16 = slice_by_index(begin = var_2933_begin_0, end = var_2933_end_0, end_mask = var_2933_end_mask_0, x = shares_9_cast_fp16)[name = string("op_2933_cast_fp16")];
            bool var_2934_transpose_x_0 = const()[name = string("op_2934_transpose_x_0"), val = bool(false)];
            bool var_2934_transpose_y_0 = const()[name = string("op_2934_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_9_cast_fp16 = transpose(perm = v_9_perm_0, x = var_2789_cast_fp16)[name = string("transpose_116")];
            tensor<fp16, [1, 8, 32, 128]> var_2934_cast_fp16 = matmul(transpose_x = var_2934_transpose_x_0, transpose_y = var_2934_transpose_y_0, x = var_2933_cast_fp16, y = v_9_cast_fp16)[name = string("op_2934_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_17_cast_fp16 = add(x = var_2928_cast_fp16, y = var_2934_cast_fp16)[name = string("mixed_17_cast_fp16")];
            tensor<int32, [4]> var_2951_begin_0 = const()[name = string("op_2951_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_2951_end_0 = const()[name = string("op_2951_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_2951_end_mask_0 = const()[name = string("op_2951_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2951_cast_fp16 = slice_by_index(begin = var_2951_begin_0, end = var_2951_end_0, end_mask = var_2951_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2951_cast_fp16")];
            tensor<int32, [4]> var_2966_begin_0 = const()[name = string("op_2966_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_2966_end_0 = const()[name = string("op_2966_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_2966_end_mask_0 = const()[name = string("op_2966_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_2966_cast_fp16 = slice_by_index(begin = var_2966_begin_0, end = var_2966_end_0, end_mask = var_2966_end_mask_0, x = mixed_17_cast_fp16)[name = string("op_2966_cast_fp16")];
            int32 var_2968 = const()[name = string("op_2968"), val = int32(1)];
            bool mixed_19_interleave_0 = const()[name = string("mixed_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_19_cast_fp16 = concat(axis = var_2968, interleave = mixed_19_interleave_0, values = (var_2951_cast_fp16, var_2966_cast_fp16))[name = string("mixed_19_cast_fp16")];
            tensor<int32, [4]> var_2972_perm_0 = const()[name = string("op_2972_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_2976 = const()[name = string("op_2976"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_2972_cast_fp16 = transpose(perm = var_2972_perm_0, x = mixed_19_cast_fp16)[name = string("transpose_115")];
            tensor<fp16, [1, 16, 2048]> input_43_cast_fp16 = reshape(shape = var_2976, x = var_2972_cast_fp16)[name = string("input_43_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_4_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(378490176))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380587392))))[name = string("model_layers_4_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_31_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_self_attn_o_proj_weight_to_fp16_quantized, x = input_43_cast_fp16)[name = string("linear_31_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_59_cast_fp16 = add(x = x_49_cast_fp16, y = linear_31_cast_fp16)[name = string("x_59_cast_fp16")];
            fp16 var_2987_promoted_to_fp16 = const()[name = string("op_2987_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_2989_cast_fp16 = pow(x = x_59_cast_fp16, y = var_2987_promoted_to_fp16)[name = string("op_2989_cast_fp16")];
            tensor<int32, [1]> var_2991_axes_0 = const()[name = string("op_2991_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_2991_keep_dims_0 = const()[name = string("op_2991_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_2991_cast_fp16 = reduce_mean(axes = var_2991_axes_0, keep_dims = var_2991_keep_dims_0, x = var_2989_cast_fp16)[name = string("op_2991_cast_fp16")];
            fp16 var_2992_to_fp16 = const()[name = string("op_2992_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_2993_cast_fp16 = add(x = var_2991_cast_fp16, y = var_2992_to_fp16)[name = string("op_2993_cast_fp16")];
            fp32 var_2994_epsilon_0 = const()[name = string("op_2994_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_2994_cast_fp16 = rsqrt(epsilon = var_2994_epsilon_0, x = var_2993_cast_fp16)[name = string("op_2994_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_2995_cast_fp16 = mul(x = x_59_cast_fp16, y = var_2994_cast_fp16)[name = string("op_2995_cast_fp16")];
            tensor<fp16, [1024]> model_layers_4_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_4_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380589504)))];
            tensor<fp16, [1, 16, 1024]> input_45_cast_fp16 = mul(x = var_2995_cast_fp16, y = model_layers_4_post_attention_layernorm_weight_to_fp16)[name = string("input_45_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(380591616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383737408))))[name = string("model_layers_4_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_32_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_gate_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_32_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3000_cast_fp16 = silu(x = linear_32_cast_fp16)[name = string("op_3000_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_4_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(383743616))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386889408))))[name = string("model_layers_4_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_33_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_4_mlp_up_proj_weight_to_fp16_quantized, x = input_45_cast_fp16)[name = string("linear_33_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_49_cast_fp16 = mul(x = var_3000_cast_fp16, y = linear_33_cast_fp16)[name = string("input_49_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_4_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(386895616))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390041408))))[name = string("model_layers_4_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_34_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_4_mlp_down_proj_weight_to_fp16_quantized, x = input_49_cast_fp16)[name = string("linear_34_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_61_cast_fp16 = add(x = x_59_cast_fp16, y = linear_34_cast_fp16)[name = string("x_61_cast_fp16")];
            fp16 var_3014_promoted_to_fp16 = const()[name = string("op_3014_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3016_cast_fp16 = pow(x = x_61_cast_fp16, y = var_3014_promoted_to_fp16)[name = string("op_3016_cast_fp16")];
            tensor<int32, [1]> var_3018_axes_0 = const()[name = string("op_3018_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3018_keep_dims_0 = const()[name = string("op_3018_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3018_cast_fp16 = reduce_mean(axes = var_3018_axes_0, keep_dims = var_3018_keep_dims_0, x = var_3016_cast_fp16)[name = string("op_3018_cast_fp16")];
            fp16 var_3019_to_fp16 = const()[name = string("op_3019_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3020_cast_fp16 = add(x = var_3018_cast_fp16, y = var_3019_to_fp16)[name = string("op_3020_cast_fp16")];
            fp32 var_3021_epsilon_0 = const()[name = string("op_3021_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3021_cast_fp16 = rsqrt(epsilon = var_3021_epsilon_0, x = var_3020_cast_fp16)[name = string("op_3021_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3022_cast_fp16 = mul(x = x_61_cast_fp16, y = var_3021_cast_fp16)[name = string("op_3022_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390043520)))];
            tensor<fp16, [1, 16, 1024]> input_51_cast_fp16 = mul(x = var_3022_cast_fp16, y = model_layers_5_input_layernorm_weight_to_fp16)[name = string("input_51_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_5_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(390045632))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392142848))))[name = string("model_layers_5_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_35_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_5_self_attn_q_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_35_cast_fp16")];
            tensor<int32, [4]> var_3031 = const()[name = string("op_3031"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3032_cast_fp16 = reshape(shape = var_3031, x = linear_35_cast_fp16)[name = string("op_3032_cast_fp16")];
            tensor<int32, [4]> x_63_perm_0 = const()[name = string("x_63_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3040_promoted_to_fp16 = const()[name = string("op_3040_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_63_cast_fp16 = transpose(perm = x_63_perm_0, x = var_3032_cast_fp16)[name = string("transpose_114")];
            tensor<fp16, [1, 16, 16, 128]> var_3042_cast_fp16 = pow(x = x_63_cast_fp16, y = var_3040_promoted_to_fp16)[name = string("op_3042_cast_fp16")];
            tensor<int32, [1]> var_3044_axes_0 = const()[name = string("op_3044_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3044_keep_dims_0 = const()[name = string("op_3044_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3044_cast_fp16 = reduce_mean(axes = var_3044_axes_0, keep_dims = var_3044_keep_dims_0, x = var_3042_cast_fp16)[name = string("op_3044_cast_fp16")];
            fp16 var_3045_to_fp16 = const()[name = string("op_3045_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3046_cast_fp16 = add(x = var_3044_cast_fp16, y = var_3045_to_fp16)[name = string("op_3046_cast_fp16")];
            fp32 var_3047_epsilon_0 = const()[name = string("op_3047_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3047_cast_fp16 = rsqrt(epsilon = var_3047_epsilon_0, x = var_3046_cast_fp16)[name = string("op_3047_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3048_cast_fp16 = mul(x = x_63_cast_fp16, y = var_3047_cast_fp16)[name = string("op_3048_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147008)))];
            tensor<fp16, [1, 16, 16, 128]> x_67_cast_fp16 = mul(x = var_3048_cast_fp16, y = model_layers_5_self_attn_q_norm_weight_to_fp16)[name = string("x_67_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(392147328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393195968))))[name = string("model_layers_5_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_36_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_k_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_36_cast_fp16")];
            tensor<int32, [4]> var_3057 = const()[name = string("op_3057"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3058_cast_fp16 = reshape(shape = var_3057, x = linear_36_cast_fp16)[name = string("op_3058_cast_fp16")];
            tensor<int32, [4]> x_65_perm_0 = const()[name = string("x_65_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3066_promoted_to_fp16 = const()[name = string("op_3066_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_65_cast_fp16 = transpose(perm = x_65_perm_0, x = var_3058_cast_fp16)[name = string("transpose_113")];
            tensor<fp16, [1, 8, 16, 128]> var_3068_cast_fp16 = pow(x = x_65_cast_fp16, y = var_3066_promoted_to_fp16)[name = string("op_3068_cast_fp16")];
            tensor<int32, [1]> var_3070_axes_0 = const()[name = string("op_3070_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3070_keep_dims_0 = const()[name = string("op_3070_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3070_cast_fp16 = reduce_mean(axes = var_3070_axes_0, keep_dims = var_3070_keep_dims_0, x = var_3068_cast_fp16)[name = string("op_3070_cast_fp16")];
            fp16 var_3071_to_fp16 = const()[name = string("op_3071_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3072_cast_fp16 = add(x = var_3070_cast_fp16, y = var_3071_to_fp16)[name = string("op_3072_cast_fp16")];
            fp32 var_3073_epsilon_0 = const()[name = string("op_3073_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3073_cast_fp16 = rsqrt(epsilon = var_3073_epsilon_0, x = var_3072_cast_fp16)[name = string("op_3073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3074_cast_fp16 = mul(x = x_65_cast_fp16, y = var_3073_cast_fp16)[name = string("op_3074_cast_fp16")];
            tensor<fp16, [128]> model_layers_5_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_5_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198080)))];
            tensor<fp16, [1, 8, 16, 128]> x_69_cast_fp16 = mul(x = var_3074_cast_fp16, y = model_layers_5_self_attn_k_norm_weight_to_fp16)[name = string("x_69_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_5_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(393198400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394247040))))[name = string("model_layers_5_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_37_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_v_proj_weight_to_fp16_quantized, x = input_51_cast_fp16)[name = string("linear_37_cast_fp16")];
            tensor<int32, [4]> var_3083 = const()[name = string("op_3083"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3084_cast_fp16 = reshape(shape = var_3083, x = linear_37_cast_fp16)[name = string("op_3084_cast_fp16")];
            tensor<int32, [4]> v_11_perm_0 = const()[name = string("v_11_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_21_begin_0 = const()[name = string("first_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_21_end_0 = const()[name = string("first_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_21_end_mask_0 = const()[name = string("first_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_21_cast_fp16 = slice_by_index(begin = first_21_begin_0, end = first_21_end_0, end_mask = first_21_end_mask_0, x = x_67_cast_fp16)[name = string("first_21_cast_fp16")];
            tensor<int32, [4]> second_21_begin_0 = const()[name = string("second_21_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_21_end_0 = const()[name = string("second_21_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_21_end_mask_0 = const()[name = string("second_21_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_21_cast_fp16 = slice_by_index(begin = second_21_begin_0, end = second_21_end_0, end_mask = second_21_end_mask_0, x = x_67_cast_fp16)[name = string("second_21_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3098_cast_fp16 = mul(x = first_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3098_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3099_cast_fp16 = mul(x = second_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3099_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3101_cast_fp16 = sub(x = var_3098_cast_fp16, y = var_3099_cast_fp16)[name = string("op_3101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3102_cast_fp16 = mul(x = second_21_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3102_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3103_cast_fp16 = mul(x = first_21_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3103_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3105_cast_fp16 = add(x = var_3102_cast_fp16, y = var_3103_cast_fp16)[name = string("op_3105_cast_fp16")];
            int32 var_3107 = const()[name = string("op_3107"), val = int32(-1)];
            bool q_11_interleave_0 = const()[name = string("q_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_11_cast_fp16 = concat(axis = var_3107, interleave = q_11_interleave_0, values = (var_3101_cast_fp16, var_3105_cast_fp16))[name = string("q_11_cast_fp16")];
            tensor<int32, [4]> first_23_begin_0 = const()[name = string("first_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_23_end_0 = const()[name = string("first_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_23_end_mask_0 = const()[name = string("first_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_23_cast_fp16 = slice_by_index(begin = first_23_begin_0, end = first_23_end_0, end_mask = first_23_end_mask_0, x = x_69_cast_fp16)[name = string("first_23_cast_fp16")];
            tensor<int32, [4]> second_23_begin_0 = const()[name = string("second_23_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_23_end_0 = const()[name = string("second_23_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_23_end_mask_0 = const()[name = string("second_23_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_23_cast_fp16 = slice_by_index(begin = second_23_begin_0, end = second_23_end_0, end_mask = second_23_end_mask_0, x = x_69_cast_fp16)[name = string("second_23_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3119_cast_fp16 = mul(x = first_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3119_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3120_cast_fp16 = mul(x = second_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3120_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3122_cast_fp16 = sub(x = var_3119_cast_fp16, y = var_3120_cast_fp16)[name = string("op_3122_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3123_cast_fp16 = mul(x = second_23_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3123_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3124_cast_fp16 = mul(x = first_23_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3124_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3126_cast_fp16 = add(x = var_3123_cast_fp16, y = var_3124_cast_fp16)[name = string("op_3126_cast_fp16")];
            int32 var_3128 = const()[name = string("op_3128"), val = int32(-1)];
            bool k_11_interleave_0 = const()[name = string("k_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_11_cast_fp16 = concat(axis = var_3128, interleave = k_11_interleave_0, values = (var_3122_cast_fp16, var_3126_cast_fp16))[name = string("k_11_cast_fp16")];
            tensor<int32, [4]> var_3139_begin_0 = const()[name = string("op_3139_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3139_end_0 = const()[name = string("op_3139_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3139_end_mask_0 = const()[name = string("op_3139_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3139_stride_0 = const()[name = string("op_3139_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3139_cast_fp16 = slice_by_index(begin = var_3139_begin_0, end = var_3139_end_0, end_mask = var_3139_end_mask_0, stride = var_3139_stride_0, x = q_11_cast_fp16)[name = string("op_3139_cast_fp16")];
            tensor<int32, [4]> var_3149_begin_0 = const()[name = string("op_3149_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3149_end_0 = const()[name = string("op_3149_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3149_end_mask_0 = const()[name = string("op_3149_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3149_stride_0 = const()[name = string("op_3149_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3149_cast_fp16 = slice_by_index(begin = var_3149_begin_0, end = var_3149_end_0, end_mask = var_3149_end_mask_0, stride = var_3149_stride_0, x = q_11_cast_fp16)[name = string("op_3149_cast_fp16")];
            int32 var_3151 = const()[name = string("op_3151"), val = int32(2)];
            bool stacked_11_interleave_0 = const()[name = string("stacked_11_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_11_cast_fp16 = concat(axis = var_3151, interleave = stacked_11_interleave_0, values = (var_3139_cast_fp16, var_3149_cast_fp16))[name = string("stacked_11_cast_fp16")];
            tensor<int32, [4]> var_3157_begin_0 = const()[name = string("op_3157_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3157_end_0 = const()[name = string("op_3157_end_0"), val = tensor<int32, [4]>([6, 8, 528, 128])];
            tensor<bool, [4]> var_3157_end_mask_0 = const()[name = string("op_3157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3157_cast_fp16 = slice_by_index(begin = var_3157_begin_0, end = var_3157_end_0, end_mask = var_3157_end_mask_0, x = read_state_0)[name = string("op_3157_cast_fp16")];
            tensor<int32, [4]> var_3167_begin_0 = const()[name = string("op_3167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3167_end_0 = const()[name = string("op_3167_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3167_end_mask_0 = const()[name = string("op_3167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3167_cast_fp16 = slice_by_index(begin = var_3167_begin_0, end = var_3167_end_0, end_mask = var_3167_end_mask_0, x = var_3157_cast_fp16)[name = string("op_3167_cast_fp16")];
            bool var_3176_transpose_x_1 = const()[name = string("op_3176_transpose_x_1"), val = bool(false)];
            bool var_3176_transpose_y_1 = const()[name = string("op_3176_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3176_cast_fp16 = matmul(transpose_x = var_3176_transpose_x_1, transpose_y = var_3176_transpose_y_1, x = var_3167_cast_fp16, y = stacked_11_cast_fp16)[name = string("op_3176_cast_fp16")];
            tensor<int32, [4]> var_3179_perm_0 = const()[name = string("op_3179_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3180_to_fp16 = const()[name = string("op_3180_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3179_cast_fp16 = transpose(perm = var_3179_perm_0, x = var_3176_cast_fp16)[name = string("transpose_112")];
            tensor<fp16, [1, 8, 32, 512]> var_3181_cast_fp16 = mul(x = var_3179_cast_fp16, y = var_3180_to_fp16)[name = string("op_3181_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_11_cast_fp16 = add(x = var_3181_cast_fp16, y = cacheMask)[name = string("cache_weights_11_cast_fp16")];
            bool var_3187_transpose_x_1 = const()[name = string("op_3187_transpose_x_1"), val = bool(false)];
            bool var_3187_transpose_y_1 = const()[name = string("op_3187_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3187_cast_fp16 = matmul(transpose_x = var_3187_transpose_x_1, transpose_y = var_3187_transpose_y_1, x = stacked_11_cast_fp16, y = k_11_cast_fp16)[name = string("op_3187_cast_fp16")];
            fp16 var_3188_to_fp16 = const()[name = string("op_3188_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3189_cast_fp16 = mul(x = var_3187_cast_fp16, y = var_3188_to_fp16)[name = string("op_3189_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_11_cast_fp16 = add(x = var_3189_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_11_cast_fp16")];
            int32 var_3193 = const()[name = string("op_3193"), val = int32(-1)];
            bool var_3194_interleave_0 = const()[name = string("op_3194_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3194_cast_fp16 = concat(axis = var_3193, interleave = var_3194_interleave_0, values = (cache_weights_11_cast_fp16, call_weights_11_cast_fp16))[name = string("op_3194_cast_fp16")];
            int32 var_3195 = const()[name = string("op_3195"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_11_cast_fp16 = softmax(axis = var_3195, x = var_3194_cast_fp16)[name = string("shares_11_cast_fp16")];
            tensor<int32, [4]> var_3202_begin_0 = const()[name = string("op_3202_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3202_end_0 = const()[name = string("op_3202_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3202_end_mask_0 = const()[name = string("op_3202_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3202_cast_fp16 = slice_by_index(begin = var_3202_begin_0, end = var_3202_end_0, end_mask = var_3202_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3202_cast_fp16")];
            tensor<int32, [4]> var_3207_begin_0 = const()[name = string("op_3207_begin_0"), val = tensor<int32, [4]>([5, 0, 0, 0])];
            tensor<int32, [4]> var_3207_end_0 = const()[name = string("op_3207_end_0"), val = tensor<int32, [4]>([6, 8, 528, 128])];
            tensor<bool, [4]> var_3207_end_mask_0 = const()[name = string("op_3207_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3207_cast_fp16 = slice_by_index(begin = var_3207_begin_0, end = var_3207_end_0, end_mask = var_3207_end_mask_0, x = read_state_1)[name = string("op_3207_cast_fp16")];
            tensor<int32, [4]> var_3217_begin_0 = const()[name = string("op_3217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3217_end_0 = const()[name = string("op_3217_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3217_end_mask_0 = const()[name = string("op_3217_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3217_cast_fp16 = slice_by_index(begin = var_3217_begin_0, end = var_3217_end_0, end_mask = var_3217_end_mask_0, x = var_3207_cast_fp16)[name = string("op_3217_cast_fp16")];
            bool var_3223_transpose_x_0 = const()[name = string("op_3223_transpose_x_0"), val = bool(false)];
            bool var_3223_transpose_y_0 = const()[name = string("op_3223_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3223_cast_fp16 = matmul(transpose_x = var_3223_transpose_x_0, transpose_y = var_3223_transpose_y_0, x = var_3202_cast_fp16, y = var_3217_cast_fp16)[name = string("op_3223_cast_fp16")];
            tensor<int32, [4]> var_3228_begin_0 = const()[name = string("op_3228_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3228_end_0 = const()[name = string("op_3228_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3228_end_mask_0 = const()[name = string("op_3228_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3228_cast_fp16 = slice_by_index(begin = var_3228_begin_0, end = var_3228_end_0, end_mask = var_3228_end_mask_0, x = shares_11_cast_fp16)[name = string("op_3228_cast_fp16")];
            bool var_3229_transpose_x_0 = const()[name = string("op_3229_transpose_x_0"), val = bool(false)];
            bool var_3229_transpose_y_0 = const()[name = string("op_3229_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_11_cast_fp16 = transpose(perm = v_11_perm_0, x = var_3084_cast_fp16)[name = string("transpose_111")];
            tensor<fp16, [1, 8, 32, 128]> var_3229_cast_fp16 = matmul(transpose_x = var_3229_transpose_x_0, transpose_y = var_3229_transpose_y_0, x = var_3228_cast_fp16, y = v_11_cast_fp16)[name = string("op_3229_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_21_cast_fp16 = add(x = var_3223_cast_fp16, y = var_3229_cast_fp16)[name = string("mixed_21_cast_fp16")];
            tensor<int32, [4]> var_3246_begin_0 = const()[name = string("op_3246_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3246_end_0 = const()[name = string("op_3246_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3246_end_mask_0 = const()[name = string("op_3246_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3246_cast_fp16 = slice_by_index(begin = var_3246_begin_0, end = var_3246_end_0, end_mask = var_3246_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3246_cast_fp16")];
            tensor<int32, [4]> var_3261_begin_0 = const()[name = string("op_3261_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3261_end_0 = const()[name = string("op_3261_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3261_end_mask_0 = const()[name = string("op_3261_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3261_cast_fp16 = slice_by_index(begin = var_3261_begin_0, end = var_3261_end_0, end_mask = var_3261_end_mask_0, x = mixed_21_cast_fp16)[name = string("op_3261_cast_fp16")];
            int32 var_3263 = const()[name = string("op_3263"), val = int32(1)];
            bool mixed_23_interleave_0 = const()[name = string("mixed_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_23_cast_fp16 = concat(axis = var_3263, interleave = mixed_23_interleave_0, values = (var_3246_cast_fp16, var_3261_cast_fp16))[name = string("mixed_23_cast_fp16")];
            tensor<int32, [4]> var_3267_perm_0 = const()[name = string("op_3267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3271 = const()[name = string("op_3271"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3267_cast_fp16 = transpose(perm = var_3267_perm_0, x = mixed_23_cast_fp16)[name = string("transpose_110")];
            tensor<fp16, [1, 16, 2048]> input_53_cast_fp16 = reshape(shape = var_3271, x = var_3267_cast_fp16)[name = string("input_53_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_5_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(394249152))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396346368))))[name = string("model_layers_5_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_38_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_self_attn_o_proj_weight_to_fp16_quantized, x = input_53_cast_fp16)[name = string("linear_38_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_71_cast_fp16 = add(x = x_61_cast_fp16, y = linear_38_cast_fp16)[name = string("x_71_cast_fp16")];
            fp16 var_3282_promoted_to_fp16 = const()[name = string("op_3282_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3284_cast_fp16 = pow(x = x_71_cast_fp16, y = var_3282_promoted_to_fp16)[name = string("op_3284_cast_fp16")];
            tensor<int32, [1]> var_3286_axes_0 = const()[name = string("op_3286_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3286_keep_dims_0 = const()[name = string("op_3286_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3286_cast_fp16 = reduce_mean(axes = var_3286_axes_0, keep_dims = var_3286_keep_dims_0, x = var_3284_cast_fp16)[name = string("op_3286_cast_fp16")];
            fp16 var_3287_to_fp16 = const()[name = string("op_3287_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3288_cast_fp16 = add(x = var_3286_cast_fp16, y = var_3287_to_fp16)[name = string("op_3288_cast_fp16")];
            fp32 var_3289_epsilon_0 = const()[name = string("op_3289_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3289_cast_fp16 = rsqrt(epsilon = var_3289_epsilon_0, x = var_3288_cast_fp16)[name = string("op_3289_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3290_cast_fp16 = mul(x = x_71_cast_fp16, y = var_3289_cast_fp16)[name = string("op_3290_cast_fp16")];
            tensor<fp16, [1024]> model_layers_5_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_5_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396348480)))];
            tensor<fp16, [1, 16, 1024]> input_55_cast_fp16 = mul(x = var_3290_cast_fp16, y = model_layers_5_post_attention_layernorm_weight_to_fp16)[name = string("input_55_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(396350592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399496384))))[name = string("model_layers_5_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_39_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_gate_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_39_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3295_cast_fp16 = silu(x = linear_39_cast_fp16)[name = string("op_3295_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_5_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(399502592))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402648384))))[name = string("model_layers_5_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_40_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_5_mlp_up_proj_weight_to_fp16_quantized, x = input_55_cast_fp16)[name = string("linear_40_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_59_cast_fp16 = mul(x = var_3295_cast_fp16, y = linear_40_cast_fp16)[name = string("input_59_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_5_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(402654592))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405800384))))[name = string("model_layers_5_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_41_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_5_mlp_down_proj_weight_to_fp16_quantized, x = input_59_cast_fp16)[name = string("linear_41_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_73_cast_fp16 = add(x = x_71_cast_fp16, y = linear_41_cast_fp16)[name = string("x_73_cast_fp16")];
            fp16 var_3309_promoted_to_fp16 = const()[name = string("op_3309_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3311_cast_fp16 = pow(x = x_73_cast_fp16, y = var_3309_promoted_to_fp16)[name = string("op_3311_cast_fp16")];
            tensor<int32, [1]> var_3313_axes_0 = const()[name = string("op_3313_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3313_keep_dims_0 = const()[name = string("op_3313_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3313_cast_fp16 = reduce_mean(axes = var_3313_axes_0, keep_dims = var_3313_keep_dims_0, x = var_3311_cast_fp16)[name = string("op_3313_cast_fp16")];
            fp16 var_3314_to_fp16 = const()[name = string("op_3314_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3315_cast_fp16 = add(x = var_3313_cast_fp16, y = var_3314_to_fp16)[name = string("op_3315_cast_fp16")];
            fp32 var_3316_epsilon_0 = const()[name = string("op_3316_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3316_cast_fp16 = rsqrt(epsilon = var_3316_epsilon_0, x = var_3315_cast_fp16)[name = string("op_3316_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3317_cast_fp16 = mul(x = x_73_cast_fp16, y = var_3316_cast_fp16)[name = string("op_3317_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405802496)))];
            tensor<fp16, [1, 16, 1024]> input_61_cast_fp16 = mul(x = var_3317_cast_fp16, y = model_layers_6_input_layernorm_weight_to_fp16)[name = string("input_61_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_6_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(405804608))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407901824))))[name = string("model_layers_6_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_42_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_6_self_attn_q_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_42_cast_fp16")];
            tensor<int32, [4]> var_3326 = const()[name = string("op_3326"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3327_cast_fp16 = reshape(shape = var_3326, x = linear_42_cast_fp16)[name = string("op_3327_cast_fp16")];
            tensor<int32, [4]> x_75_perm_0 = const()[name = string("x_75_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3335_promoted_to_fp16 = const()[name = string("op_3335_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_75_cast_fp16 = transpose(perm = x_75_perm_0, x = var_3327_cast_fp16)[name = string("transpose_109")];
            tensor<fp16, [1, 16, 16, 128]> var_3337_cast_fp16 = pow(x = x_75_cast_fp16, y = var_3335_promoted_to_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]>([-1])];
            bool var_3339_keep_dims_0 = const()[name = string("op_3339_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3339_cast_fp16 = reduce_mean(axes = var_3339_axes_0, keep_dims = var_3339_keep_dims_0, x = var_3337_cast_fp16)[name = string("op_3339_cast_fp16")];
            fp16 var_3340_to_fp16 = const()[name = string("op_3340_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3341_cast_fp16 = add(x = var_3339_cast_fp16, y = var_3340_to_fp16)[name = string("op_3341_cast_fp16")];
            fp32 var_3342_epsilon_0 = const()[name = string("op_3342_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3342_cast_fp16 = rsqrt(epsilon = var_3342_epsilon_0, x = var_3341_cast_fp16)[name = string("op_3342_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3343_cast_fp16 = mul(x = x_75_cast_fp16, y = var_3342_cast_fp16)[name = string("op_3343_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407905984)))];
            tensor<fp16, [1, 16, 16, 128]> x_79_cast_fp16 = mul(x = var_3343_cast_fp16, y = model_layers_6_self_attn_q_norm_weight_to_fp16)[name = string("x_79_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(407906304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408954944))))[name = string("model_layers_6_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_43_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_k_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_43_cast_fp16")];
            tensor<int32, [4]> var_3352 = const()[name = string("op_3352"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3353_cast_fp16 = reshape(shape = var_3352, x = linear_43_cast_fp16)[name = string("op_3353_cast_fp16")];
            tensor<int32, [4]> x_77_perm_0 = const()[name = string("x_77_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3361_promoted_to_fp16 = const()[name = string("op_3361_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_77_cast_fp16 = transpose(perm = x_77_perm_0, x = var_3353_cast_fp16)[name = string("transpose_108")];
            tensor<fp16, [1, 8, 16, 128]> var_3363_cast_fp16 = pow(x = x_77_cast_fp16, y = var_3361_promoted_to_fp16)[name = string("op_3363_cast_fp16")];
            tensor<int32, [1]> var_3365_axes_0 = const()[name = string("op_3365_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3365_keep_dims_0 = const()[name = string("op_3365_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3365_cast_fp16 = reduce_mean(axes = var_3365_axes_0, keep_dims = var_3365_keep_dims_0, x = var_3363_cast_fp16)[name = string("op_3365_cast_fp16")];
            fp16 var_3366_to_fp16 = const()[name = string("op_3366_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3367_cast_fp16 = add(x = var_3365_cast_fp16, y = var_3366_to_fp16)[name = string("op_3367_cast_fp16")];
            fp32 var_3368_epsilon_0 = const()[name = string("op_3368_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3368_cast_fp16 = rsqrt(epsilon = var_3368_epsilon_0, x = var_3367_cast_fp16)[name = string("op_3368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3369_cast_fp16 = mul(x = x_77_cast_fp16, y = var_3368_cast_fp16)[name = string("op_3369_cast_fp16")];
            tensor<fp16, [128]> model_layers_6_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_6_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957056)))];
            tensor<fp16, [1, 8, 16, 128]> x_81_cast_fp16 = mul(x = var_3369_cast_fp16, y = model_layers_6_self_attn_k_norm_weight_to_fp16)[name = string("x_81_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_6_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(408957376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410006016))))[name = string("model_layers_6_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_44_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_v_proj_weight_to_fp16_quantized, x = input_61_cast_fp16)[name = string("linear_44_cast_fp16")];
            tensor<int32, [4]> var_3378 = const()[name = string("op_3378"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3379_cast_fp16 = reshape(shape = var_3378, x = linear_44_cast_fp16)[name = string("op_3379_cast_fp16")];
            tensor<int32, [4]> v_13_perm_0 = const()[name = string("v_13_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_25_begin_0 = const()[name = string("first_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_25_end_0 = const()[name = string("first_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_25_end_mask_0 = const()[name = string("first_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_25_cast_fp16 = slice_by_index(begin = first_25_begin_0, end = first_25_end_0, end_mask = first_25_end_mask_0, x = x_79_cast_fp16)[name = string("first_25_cast_fp16")];
            tensor<int32, [4]> second_25_begin_0 = const()[name = string("second_25_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_25_end_0 = const()[name = string("second_25_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_25_end_mask_0 = const()[name = string("second_25_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_25_cast_fp16 = slice_by_index(begin = second_25_begin_0, end = second_25_end_0, end_mask = second_25_end_mask_0, x = x_79_cast_fp16)[name = string("second_25_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3393_cast_fp16 = mul(x = first_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3393_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3394_cast_fp16 = mul(x = second_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3394_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3396_cast_fp16 = sub(x = var_3393_cast_fp16, y = var_3394_cast_fp16)[name = string("op_3396_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3397_cast_fp16 = mul(x = second_25_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3397_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3398_cast_fp16 = mul(x = first_25_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3398_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3400_cast_fp16 = add(x = var_3397_cast_fp16, y = var_3398_cast_fp16)[name = string("op_3400_cast_fp16")];
            int32 var_3402 = const()[name = string("op_3402"), val = int32(-1)];
            bool q_13_interleave_0 = const()[name = string("q_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_13_cast_fp16 = concat(axis = var_3402, interleave = q_13_interleave_0, values = (var_3396_cast_fp16, var_3400_cast_fp16))[name = string("q_13_cast_fp16")];
            tensor<int32, [4]> first_27_begin_0 = const()[name = string("first_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_27_end_0 = const()[name = string("first_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_27_end_mask_0 = const()[name = string("first_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_27_cast_fp16 = slice_by_index(begin = first_27_begin_0, end = first_27_end_0, end_mask = first_27_end_mask_0, x = x_81_cast_fp16)[name = string("first_27_cast_fp16")];
            tensor<int32, [4]> second_27_begin_0 = const()[name = string("second_27_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_27_end_0 = const()[name = string("second_27_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_27_end_mask_0 = const()[name = string("second_27_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_27_cast_fp16 = slice_by_index(begin = second_27_begin_0, end = second_27_end_0, end_mask = second_27_end_mask_0, x = x_81_cast_fp16)[name = string("second_27_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3414_cast_fp16 = mul(x = first_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3414_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3415_cast_fp16 = mul(x = second_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3415_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3417_cast_fp16 = sub(x = var_3414_cast_fp16, y = var_3415_cast_fp16)[name = string("op_3417_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3418_cast_fp16 = mul(x = second_27_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3418_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3419_cast_fp16 = mul(x = first_27_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3419_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3421_cast_fp16 = add(x = var_3418_cast_fp16, y = var_3419_cast_fp16)[name = string("op_3421_cast_fp16")];
            int32 var_3423 = const()[name = string("op_3423"), val = int32(-1)];
            bool k_13_interleave_0 = const()[name = string("k_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_13_cast_fp16 = concat(axis = var_3423, interleave = k_13_interleave_0, values = (var_3417_cast_fp16, var_3421_cast_fp16))[name = string("k_13_cast_fp16")];
            tensor<int32, [4]> var_3434_begin_0 = const()[name = string("op_3434_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3434_end_0 = const()[name = string("op_3434_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3434_end_mask_0 = const()[name = string("op_3434_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3434_stride_0 = const()[name = string("op_3434_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3434_cast_fp16 = slice_by_index(begin = var_3434_begin_0, end = var_3434_end_0, end_mask = var_3434_end_mask_0, stride = var_3434_stride_0, x = q_13_cast_fp16)[name = string("op_3434_cast_fp16")];
            tensor<int32, [4]> var_3444_begin_0 = const()[name = string("op_3444_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3444_end_0 = const()[name = string("op_3444_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3444_end_mask_0 = const()[name = string("op_3444_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3444_stride_0 = const()[name = string("op_3444_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3444_cast_fp16 = slice_by_index(begin = var_3444_begin_0, end = var_3444_end_0, end_mask = var_3444_end_mask_0, stride = var_3444_stride_0, x = q_13_cast_fp16)[name = string("op_3444_cast_fp16")];
            int32 var_3446 = const()[name = string("op_3446"), val = int32(2)];
            bool stacked_13_interleave_0 = const()[name = string("stacked_13_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_13_cast_fp16 = concat(axis = var_3446, interleave = stacked_13_interleave_0, values = (var_3434_cast_fp16, var_3444_cast_fp16))[name = string("stacked_13_cast_fp16")];
            tensor<int32, [4]> var_3452_begin_0 = const()[name = string("op_3452_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3452_end_0 = const()[name = string("op_3452_end_0"), val = tensor<int32, [4]>([7, 8, 528, 128])];
            tensor<bool, [4]> var_3452_end_mask_0 = const()[name = string("op_3452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3452_cast_fp16 = slice_by_index(begin = var_3452_begin_0, end = var_3452_end_0, end_mask = var_3452_end_mask_0, x = read_state_0)[name = string("op_3452_cast_fp16")];
            tensor<int32, [4]> var_3462_begin_0 = const()[name = string("op_3462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3462_end_0 = const()[name = string("op_3462_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3462_end_mask_0 = const()[name = string("op_3462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3462_cast_fp16 = slice_by_index(begin = var_3462_begin_0, end = var_3462_end_0, end_mask = var_3462_end_mask_0, x = var_3452_cast_fp16)[name = string("op_3462_cast_fp16")];
            bool var_3471_transpose_x_1 = const()[name = string("op_3471_transpose_x_1"), val = bool(false)];
            bool var_3471_transpose_y_1 = const()[name = string("op_3471_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3471_cast_fp16 = matmul(transpose_x = var_3471_transpose_x_1, transpose_y = var_3471_transpose_y_1, x = var_3462_cast_fp16, y = stacked_13_cast_fp16)[name = string("op_3471_cast_fp16")];
            tensor<int32, [4]> var_3474_perm_0 = const()[name = string("op_3474_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3475_to_fp16 = const()[name = string("op_3475_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3474_cast_fp16 = transpose(perm = var_3474_perm_0, x = var_3471_cast_fp16)[name = string("transpose_107")];
            tensor<fp16, [1, 8, 32, 512]> var_3476_cast_fp16 = mul(x = var_3474_cast_fp16, y = var_3475_to_fp16)[name = string("op_3476_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_13_cast_fp16 = add(x = var_3476_cast_fp16, y = cacheMask)[name = string("cache_weights_13_cast_fp16")];
            bool var_3482_transpose_x_1 = const()[name = string("op_3482_transpose_x_1"), val = bool(false)];
            bool var_3482_transpose_y_1 = const()[name = string("op_3482_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3482_cast_fp16 = matmul(transpose_x = var_3482_transpose_x_1, transpose_y = var_3482_transpose_y_1, x = stacked_13_cast_fp16, y = k_13_cast_fp16)[name = string("op_3482_cast_fp16")];
            fp16 var_3483_to_fp16 = const()[name = string("op_3483_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3484_cast_fp16 = mul(x = var_3482_cast_fp16, y = var_3483_to_fp16)[name = string("op_3484_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_13_cast_fp16 = add(x = var_3484_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_13_cast_fp16")];
            int32 var_3488 = const()[name = string("op_3488"), val = int32(-1)];
            bool var_3489_interleave_0 = const()[name = string("op_3489_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3489_cast_fp16 = concat(axis = var_3488, interleave = var_3489_interleave_0, values = (cache_weights_13_cast_fp16, call_weights_13_cast_fp16))[name = string("op_3489_cast_fp16")];
            int32 var_3490 = const()[name = string("op_3490"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_13_cast_fp16 = softmax(axis = var_3490, x = var_3489_cast_fp16)[name = string("shares_13_cast_fp16")];
            tensor<int32, [4]> var_3497_begin_0 = const()[name = string("op_3497_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3497_end_0 = const()[name = string("op_3497_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3497_end_mask_0 = const()[name = string("op_3497_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3497_cast_fp16 = slice_by_index(begin = var_3497_begin_0, end = var_3497_end_0, end_mask = var_3497_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3497_cast_fp16")];
            tensor<int32, [4]> var_3502_begin_0 = const()[name = string("op_3502_begin_0"), val = tensor<int32, [4]>([6, 0, 0, 0])];
            tensor<int32, [4]> var_3502_end_0 = const()[name = string("op_3502_end_0"), val = tensor<int32, [4]>([7, 8, 528, 128])];
            tensor<bool, [4]> var_3502_end_mask_0 = const()[name = string("op_3502_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3502_cast_fp16 = slice_by_index(begin = var_3502_begin_0, end = var_3502_end_0, end_mask = var_3502_end_mask_0, x = read_state_1)[name = string("op_3502_cast_fp16")];
            tensor<int32, [4]> var_3512_begin_0 = const()[name = string("op_3512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3512_end_0 = const()[name = string("op_3512_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3512_end_mask_0 = const()[name = string("op_3512_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3512_cast_fp16 = slice_by_index(begin = var_3512_begin_0, end = var_3512_end_0, end_mask = var_3512_end_mask_0, x = var_3502_cast_fp16)[name = string("op_3512_cast_fp16")];
            bool var_3518_transpose_x_0 = const()[name = string("op_3518_transpose_x_0"), val = bool(false)];
            bool var_3518_transpose_y_0 = const()[name = string("op_3518_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3518_cast_fp16 = matmul(transpose_x = var_3518_transpose_x_0, transpose_y = var_3518_transpose_y_0, x = var_3497_cast_fp16, y = var_3512_cast_fp16)[name = string("op_3518_cast_fp16")];
            tensor<int32, [4]> var_3523_begin_0 = const()[name = string("op_3523_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3523_end_0 = const()[name = string("op_3523_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3523_end_mask_0 = const()[name = string("op_3523_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3523_cast_fp16 = slice_by_index(begin = var_3523_begin_0, end = var_3523_end_0, end_mask = var_3523_end_mask_0, x = shares_13_cast_fp16)[name = string("op_3523_cast_fp16")];
            bool var_3524_transpose_x_0 = const()[name = string("op_3524_transpose_x_0"), val = bool(false)];
            bool var_3524_transpose_y_0 = const()[name = string("op_3524_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_13_cast_fp16 = transpose(perm = v_13_perm_0, x = var_3379_cast_fp16)[name = string("transpose_106")];
            tensor<fp16, [1, 8, 32, 128]> var_3524_cast_fp16 = matmul(transpose_x = var_3524_transpose_x_0, transpose_y = var_3524_transpose_y_0, x = var_3523_cast_fp16, y = v_13_cast_fp16)[name = string("op_3524_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_25_cast_fp16 = add(x = var_3518_cast_fp16, y = var_3524_cast_fp16)[name = string("mixed_25_cast_fp16")];
            tensor<int32, [4]> var_3541_begin_0 = const()[name = string("op_3541_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3541_end_0 = const()[name = string("op_3541_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3541_end_mask_0 = const()[name = string("op_3541_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3541_cast_fp16 = slice_by_index(begin = var_3541_begin_0, end = var_3541_end_0, end_mask = var_3541_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3541_cast_fp16")];
            tensor<int32, [4]> var_3556_begin_0 = const()[name = string("op_3556_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3556_end_0 = const()[name = string("op_3556_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3556_end_mask_0 = const()[name = string("op_3556_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3556_cast_fp16 = slice_by_index(begin = var_3556_begin_0, end = var_3556_end_0, end_mask = var_3556_end_mask_0, x = mixed_25_cast_fp16)[name = string("op_3556_cast_fp16")];
            int32 var_3558 = const()[name = string("op_3558"), val = int32(1)];
            bool mixed_27_interleave_0 = const()[name = string("mixed_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_27_cast_fp16 = concat(axis = var_3558, interleave = mixed_27_interleave_0, values = (var_3541_cast_fp16, var_3556_cast_fp16))[name = string("mixed_27_cast_fp16")];
            tensor<int32, [4]> var_3562_perm_0 = const()[name = string("op_3562_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3566 = const()[name = string("op_3566"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3562_cast_fp16 = transpose(perm = var_3562_perm_0, x = mixed_27_cast_fp16)[name = string("transpose_105")];
            tensor<fp16, [1, 16, 2048]> input_63_cast_fp16 = reshape(shape = var_3566, x = var_3562_cast_fp16)[name = string("input_63_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_6_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(410008128))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412105344))))[name = string("model_layers_6_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_45_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_self_attn_o_proj_weight_to_fp16_quantized, x = input_63_cast_fp16)[name = string("linear_45_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_83_cast_fp16 = add(x = x_73_cast_fp16, y = linear_45_cast_fp16)[name = string("x_83_cast_fp16")];
            fp16 var_3577_promoted_to_fp16 = const()[name = string("op_3577_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3579_cast_fp16 = pow(x = x_83_cast_fp16, y = var_3577_promoted_to_fp16)[name = string("op_3579_cast_fp16")];
            tensor<int32, [1]> var_3581_axes_0 = const()[name = string("op_3581_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3581_keep_dims_0 = const()[name = string("op_3581_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3581_cast_fp16 = reduce_mean(axes = var_3581_axes_0, keep_dims = var_3581_keep_dims_0, x = var_3579_cast_fp16)[name = string("op_3581_cast_fp16")];
            fp16 var_3582_to_fp16 = const()[name = string("op_3582_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3583_cast_fp16 = add(x = var_3581_cast_fp16, y = var_3582_to_fp16)[name = string("op_3583_cast_fp16")];
            fp32 var_3584_epsilon_0 = const()[name = string("op_3584_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3584_cast_fp16 = rsqrt(epsilon = var_3584_epsilon_0, x = var_3583_cast_fp16)[name = string("op_3584_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3585_cast_fp16 = mul(x = x_83_cast_fp16, y = var_3584_cast_fp16)[name = string("op_3585_cast_fp16")];
            tensor<fp16, [1024]> model_layers_6_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_6_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412107456)))];
            tensor<fp16, [1, 16, 1024]> input_65_cast_fp16 = mul(x = var_3585_cast_fp16, y = model_layers_6_post_attention_layernorm_weight_to_fp16)[name = string("input_65_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(412109568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415255360))))[name = string("model_layers_6_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_46_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_gate_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_46_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3590_cast_fp16 = silu(x = linear_46_cast_fp16)[name = string("op_3590_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_6_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(415261568))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418407360))))[name = string("model_layers_6_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_47_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_6_mlp_up_proj_weight_to_fp16_quantized, x = input_65_cast_fp16)[name = string("linear_47_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_69_cast_fp16 = mul(x = var_3590_cast_fp16, y = linear_47_cast_fp16)[name = string("input_69_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_6_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(418413568))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421559360))))[name = string("model_layers_6_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_48_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_6_mlp_down_proj_weight_to_fp16_quantized, x = input_69_cast_fp16)[name = string("linear_48_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_85_cast_fp16 = add(x = x_83_cast_fp16, y = linear_48_cast_fp16)[name = string("x_85_cast_fp16")];
            fp16 var_3604_promoted_to_fp16 = const()[name = string("op_3604_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3606_cast_fp16 = pow(x = x_85_cast_fp16, y = var_3604_promoted_to_fp16)[name = string("op_3606_cast_fp16")];
            tensor<int32, [1]> var_3608_axes_0 = const()[name = string("op_3608_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3608_keep_dims_0 = const()[name = string("op_3608_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3608_cast_fp16 = reduce_mean(axes = var_3608_axes_0, keep_dims = var_3608_keep_dims_0, x = var_3606_cast_fp16)[name = string("op_3608_cast_fp16")];
            fp16 var_3609_to_fp16 = const()[name = string("op_3609_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3610_cast_fp16 = add(x = var_3608_cast_fp16, y = var_3609_to_fp16)[name = string("op_3610_cast_fp16")];
            fp32 var_3611_epsilon_0 = const()[name = string("op_3611_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3611_cast_fp16 = rsqrt(epsilon = var_3611_epsilon_0, x = var_3610_cast_fp16)[name = string("op_3611_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3612_cast_fp16 = mul(x = x_85_cast_fp16, y = var_3611_cast_fp16)[name = string("op_3612_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421561472)))];
            tensor<fp16, [1, 16, 1024]> input_71_cast_fp16 = mul(x = var_3612_cast_fp16, y = model_layers_7_input_layernorm_weight_to_fp16)[name = string("input_71_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_7_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(421563584))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423660800))))[name = string("model_layers_7_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_49_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_7_self_attn_q_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_49_cast_fp16")];
            tensor<int32, [4]> var_3621 = const()[name = string("op_3621"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3622_cast_fp16 = reshape(shape = var_3621, x = linear_49_cast_fp16)[name = string("op_3622_cast_fp16")];
            tensor<int32, [4]> x_87_perm_0 = const()[name = string("x_87_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3630_promoted_to_fp16 = const()[name = string("op_3630_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_87_cast_fp16 = transpose(perm = x_87_perm_0, x = var_3622_cast_fp16)[name = string("transpose_104")];
            tensor<fp16, [1, 16, 16, 128]> var_3632_cast_fp16 = pow(x = x_87_cast_fp16, y = var_3630_promoted_to_fp16)[name = string("op_3632_cast_fp16")];
            tensor<int32, [1]> var_3634_axes_0 = const()[name = string("op_3634_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3634_keep_dims_0 = const()[name = string("op_3634_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3634_cast_fp16 = reduce_mean(axes = var_3634_axes_0, keep_dims = var_3634_keep_dims_0, x = var_3632_cast_fp16)[name = string("op_3634_cast_fp16")];
            fp16 var_3635_to_fp16 = const()[name = string("op_3635_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3636_cast_fp16 = add(x = var_3634_cast_fp16, y = var_3635_to_fp16)[name = string("op_3636_cast_fp16")];
            fp32 var_3637_epsilon_0 = const()[name = string("op_3637_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3637_cast_fp16 = rsqrt(epsilon = var_3637_epsilon_0, x = var_3636_cast_fp16)[name = string("op_3637_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3638_cast_fp16 = mul(x = x_87_cast_fp16, y = var_3637_cast_fp16)[name = string("op_3638_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423664960)))];
            tensor<fp16, [1, 16, 16, 128]> x_91_cast_fp16 = mul(x = var_3638_cast_fp16, y = model_layers_7_self_attn_q_norm_weight_to_fp16)[name = string("x_91_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423665280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424713920))))[name = string("model_layers_7_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_50_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_k_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_50_cast_fp16")];
            tensor<int32, [4]> var_3647 = const()[name = string("op_3647"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3648_cast_fp16 = reshape(shape = var_3647, x = linear_50_cast_fp16)[name = string("op_3648_cast_fp16")];
            tensor<int32, [4]> x_89_perm_0 = const()[name = string("x_89_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3656_promoted_to_fp16 = const()[name = string("op_3656_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_89_cast_fp16 = transpose(perm = x_89_perm_0, x = var_3648_cast_fp16)[name = string("transpose_103")];
            tensor<fp16, [1, 8, 16, 128]> var_3658_cast_fp16 = pow(x = x_89_cast_fp16, y = var_3656_promoted_to_fp16)[name = string("op_3658_cast_fp16")];
            tensor<int32, [1]> var_3660_axes_0 = const()[name = string("op_3660_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3660_keep_dims_0 = const()[name = string("op_3660_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3660_cast_fp16 = reduce_mean(axes = var_3660_axes_0, keep_dims = var_3660_keep_dims_0, x = var_3658_cast_fp16)[name = string("op_3660_cast_fp16")];
            fp16 var_3661_to_fp16 = const()[name = string("op_3661_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3662_cast_fp16 = add(x = var_3660_cast_fp16, y = var_3661_to_fp16)[name = string("op_3662_cast_fp16")];
            fp32 var_3663_epsilon_0 = const()[name = string("op_3663_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3663_cast_fp16 = rsqrt(epsilon = var_3663_epsilon_0, x = var_3662_cast_fp16)[name = string("op_3663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3664_cast_fp16 = mul(x = x_89_cast_fp16, y = var_3663_cast_fp16)[name = string("op_3664_cast_fp16")];
            tensor<fp16, [128]> model_layers_7_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_7_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716032)))];
            tensor<fp16, [1, 8, 16, 128]> x_93_cast_fp16 = mul(x = var_3664_cast_fp16, y = model_layers_7_self_attn_k_norm_weight_to_fp16)[name = string("x_93_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_7_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424716352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425764992))))[name = string("model_layers_7_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_51_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_v_proj_weight_to_fp16_quantized, x = input_71_cast_fp16)[name = string("linear_51_cast_fp16")];
            tensor<int32, [4]> var_3673 = const()[name = string("op_3673"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3674_cast_fp16 = reshape(shape = var_3673, x = linear_51_cast_fp16)[name = string("op_3674_cast_fp16")];
            tensor<int32, [4]> v_15_perm_0 = const()[name = string("v_15_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_29_begin_0 = const()[name = string("first_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_29_end_0 = const()[name = string("first_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_29_end_mask_0 = const()[name = string("first_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_29_cast_fp16 = slice_by_index(begin = first_29_begin_0, end = first_29_end_0, end_mask = first_29_end_mask_0, x = x_91_cast_fp16)[name = string("first_29_cast_fp16")];
            tensor<int32, [4]> second_29_begin_0 = const()[name = string("second_29_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_29_end_0 = const()[name = string("second_29_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_29_end_mask_0 = const()[name = string("second_29_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_29_cast_fp16 = slice_by_index(begin = second_29_begin_0, end = second_29_end_0, end_mask = second_29_end_mask_0, x = x_91_cast_fp16)[name = string("second_29_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3688_cast_fp16 = mul(x = first_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3688_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3689_cast_fp16 = mul(x = second_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3689_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3691_cast_fp16 = sub(x = var_3688_cast_fp16, y = var_3689_cast_fp16)[name = string("op_3691_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3692_cast_fp16 = mul(x = second_29_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3692_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3693_cast_fp16 = mul(x = first_29_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3693_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3695_cast_fp16 = add(x = var_3692_cast_fp16, y = var_3693_cast_fp16)[name = string("op_3695_cast_fp16")];
            int32 var_3697 = const()[name = string("op_3697"), val = int32(-1)];
            bool q_15_interleave_0 = const()[name = string("q_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_15_cast_fp16 = concat(axis = var_3697, interleave = q_15_interleave_0, values = (var_3691_cast_fp16, var_3695_cast_fp16))[name = string("q_15_cast_fp16")];
            tensor<int32, [4]> first_31_begin_0 = const()[name = string("first_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_31_end_0 = const()[name = string("first_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_31_end_mask_0 = const()[name = string("first_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_31_cast_fp16 = slice_by_index(begin = first_31_begin_0, end = first_31_end_0, end_mask = first_31_end_mask_0, x = x_93_cast_fp16)[name = string("first_31_cast_fp16")];
            tensor<int32, [4]> second_31_begin_0 = const()[name = string("second_31_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_31_end_0 = const()[name = string("second_31_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_31_end_mask_0 = const()[name = string("second_31_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_31_cast_fp16 = slice_by_index(begin = second_31_begin_0, end = second_31_end_0, end_mask = second_31_end_mask_0, x = x_93_cast_fp16)[name = string("second_31_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3709_cast_fp16 = mul(x = first_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3709_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3710_cast_fp16 = mul(x = second_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3710_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3712_cast_fp16 = sub(x = var_3709_cast_fp16, y = var_3710_cast_fp16)[name = string("op_3712_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3713_cast_fp16 = mul(x = second_31_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3713_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3714_cast_fp16 = mul(x = first_31_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3714_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_3716_cast_fp16 = add(x = var_3713_cast_fp16, y = var_3714_cast_fp16)[name = string("op_3716_cast_fp16")];
            int32 var_3718 = const()[name = string("op_3718"), val = int32(-1)];
            bool k_15_interleave_0 = const()[name = string("k_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_15_cast_fp16 = concat(axis = var_3718, interleave = k_15_interleave_0, values = (var_3712_cast_fp16, var_3716_cast_fp16))[name = string("k_15_cast_fp16")];
            tensor<int32, [4]> var_3729_begin_0 = const()[name = string("op_3729_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3729_end_0 = const()[name = string("op_3729_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3729_end_mask_0 = const()[name = string("op_3729_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3729_stride_0 = const()[name = string("op_3729_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3729_cast_fp16 = slice_by_index(begin = var_3729_begin_0, end = var_3729_end_0, end_mask = var_3729_end_mask_0, stride = var_3729_stride_0, x = q_15_cast_fp16)[name = string("op_3729_cast_fp16")];
            tensor<int32, [4]> var_3739_begin_0 = const()[name = string("op_3739_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_3739_end_0 = const()[name = string("op_3739_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_3739_end_mask_0 = const()[name = string("op_3739_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_3739_stride_0 = const()[name = string("op_3739_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_3739_cast_fp16 = slice_by_index(begin = var_3739_begin_0, end = var_3739_end_0, end_mask = var_3739_end_mask_0, stride = var_3739_stride_0, x = q_15_cast_fp16)[name = string("op_3739_cast_fp16")];
            int32 var_3741 = const()[name = string("op_3741"), val = int32(2)];
            bool stacked_15_interleave_0 = const()[name = string("stacked_15_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_15_cast_fp16 = concat(axis = var_3741, interleave = stacked_15_interleave_0, values = (var_3729_cast_fp16, var_3739_cast_fp16))[name = string("stacked_15_cast_fp16")];
            tensor<int32, [4]> var_3747_begin_0 = const()[name = string("op_3747_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3747_end_0 = const()[name = string("op_3747_end_0"), val = tensor<int32, [4]>([8, 8, 528, 128])];
            tensor<bool, [4]> var_3747_end_mask_0 = const()[name = string("op_3747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3747_cast_fp16 = slice_by_index(begin = var_3747_begin_0, end = var_3747_end_0, end_mask = var_3747_end_mask_0, x = read_state_0)[name = string("op_3747_cast_fp16")];
            tensor<int32, [4]> var_3757_begin_0 = const()[name = string("op_3757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3757_end_0 = const()[name = string("op_3757_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3757_end_mask_0 = const()[name = string("op_3757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3757_cast_fp16 = slice_by_index(begin = var_3757_begin_0, end = var_3757_end_0, end_mask = var_3757_end_mask_0, x = var_3747_cast_fp16)[name = string("op_3757_cast_fp16")];
            bool var_3766_transpose_x_1 = const()[name = string("op_3766_transpose_x_1"), val = bool(false)];
            bool var_3766_transpose_y_1 = const()[name = string("op_3766_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_3766_cast_fp16 = matmul(transpose_x = var_3766_transpose_x_1, transpose_y = var_3766_transpose_y_1, x = var_3757_cast_fp16, y = stacked_15_cast_fp16)[name = string("op_3766_cast_fp16")];
            tensor<int32, [4]> var_3769_perm_0 = const()[name = string("op_3769_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_3770_to_fp16 = const()[name = string("op_3770_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_3769_cast_fp16 = transpose(perm = var_3769_perm_0, x = var_3766_cast_fp16)[name = string("transpose_102")];
            tensor<fp16, [1, 8, 32, 512]> var_3771_cast_fp16 = mul(x = var_3769_cast_fp16, y = var_3770_to_fp16)[name = string("op_3771_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_15_cast_fp16 = add(x = var_3771_cast_fp16, y = cacheMask)[name = string("cache_weights_15_cast_fp16")];
            bool var_3777_transpose_x_1 = const()[name = string("op_3777_transpose_x_1"), val = bool(false)];
            bool var_3777_transpose_y_1 = const()[name = string("op_3777_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_3777_cast_fp16 = matmul(transpose_x = var_3777_transpose_x_1, transpose_y = var_3777_transpose_y_1, x = stacked_15_cast_fp16, y = k_15_cast_fp16)[name = string("op_3777_cast_fp16")];
            fp16 var_3778_to_fp16 = const()[name = string("op_3778_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_3779_cast_fp16 = mul(x = var_3777_cast_fp16, y = var_3778_to_fp16)[name = string("op_3779_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_15_cast_fp16 = add(x = var_3779_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_15_cast_fp16")];
            int32 var_3783 = const()[name = string("op_3783"), val = int32(-1)];
            bool var_3784_interleave_0 = const()[name = string("op_3784_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_3784_cast_fp16 = concat(axis = var_3783, interleave = var_3784_interleave_0, values = (cache_weights_15_cast_fp16, call_weights_15_cast_fp16))[name = string("op_3784_cast_fp16")];
            int32 var_3785 = const()[name = string("op_3785"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_15_cast_fp16 = softmax(axis = var_3785, x = var_3784_cast_fp16)[name = string("shares_15_cast_fp16")];
            tensor<int32, [4]> var_3792_begin_0 = const()[name = string("op_3792_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3792_end_0 = const()[name = string("op_3792_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_3792_end_mask_0 = const()[name = string("op_3792_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_3792_cast_fp16 = slice_by_index(begin = var_3792_begin_0, end = var_3792_end_0, end_mask = var_3792_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3792_cast_fp16")];
            tensor<int32, [4]> var_3797_begin_0 = const()[name = string("op_3797_begin_0"), val = tensor<int32, [4]>([7, 0, 0, 0])];
            tensor<int32, [4]> var_3797_end_0 = const()[name = string("op_3797_end_0"), val = tensor<int32, [4]>([8, 8, 528, 128])];
            tensor<bool, [4]> var_3797_end_mask_0 = const()[name = string("op_3797_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_3797_cast_fp16 = slice_by_index(begin = var_3797_begin_0, end = var_3797_end_0, end_mask = var_3797_end_mask_0, x = read_state_1)[name = string("op_3797_cast_fp16")];
            tensor<int32, [4]> var_3807_begin_0 = const()[name = string("op_3807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3807_end_0 = const()[name = string("op_3807_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_3807_end_mask_0 = const()[name = string("op_3807_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_3807_cast_fp16 = slice_by_index(begin = var_3807_begin_0, end = var_3807_end_0, end_mask = var_3807_end_mask_0, x = var_3797_cast_fp16)[name = string("op_3807_cast_fp16")];
            bool var_3813_transpose_x_0 = const()[name = string("op_3813_transpose_x_0"), val = bool(false)];
            bool var_3813_transpose_y_0 = const()[name = string("op_3813_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_3813_cast_fp16 = matmul(transpose_x = var_3813_transpose_x_0, transpose_y = var_3813_transpose_y_0, x = var_3792_cast_fp16, y = var_3807_cast_fp16)[name = string("op_3813_cast_fp16")];
            tensor<int32, [4]> var_3818_begin_0 = const()[name = string("op_3818_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_3818_end_0 = const()[name = string("op_3818_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_3818_end_mask_0 = const()[name = string("op_3818_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_3818_cast_fp16 = slice_by_index(begin = var_3818_begin_0, end = var_3818_end_0, end_mask = var_3818_end_mask_0, x = shares_15_cast_fp16)[name = string("op_3818_cast_fp16")];
            bool var_3819_transpose_x_0 = const()[name = string("op_3819_transpose_x_0"), val = bool(false)];
            bool var_3819_transpose_y_0 = const()[name = string("op_3819_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_15_cast_fp16 = transpose(perm = v_15_perm_0, x = var_3674_cast_fp16)[name = string("transpose_101")];
            tensor<fp16, [1, 8, 32, 128]> var_3819_cast_fp16 = matmul(transpose_x = var_3819_transpose_x_0, transpose_y = var_3819_transpose_y_0, x = var_3818_cast_fp16, y = v_15_cast_fp16)[name = string("op_3819_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_29_cast_fp16 = add(x = var_3813_cast_fp16, y = var_3819_cast_fp16)[name = string("mixed_29_cast_fp16")];
            tensor<int32, [4]> var_3836_begin_0 = const()[name = string("op_3836_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_3836_end_0 = const()[name = string("op_3836_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_3836_end_mask_0 = const()[name = string("op_3836_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3836_cast_fp16 = slice_by_index(begin = var_3836_begin_0, end = var_3836_end_0, end_mask = var_3836_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3836_cast_fp16")];
            tensor<int32, [4]> var_3851_begin_0 = const()[name = string("op_3851_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_3851_end_0 = const()[name = string("op_3851_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_3851_end_mask_0 = const()[name = string("op_3851_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_3851_cast_fp16 = slice_by_index(begin = var_3851_begin_0, end = var_3851_end_0, end_mask = var_3851_end_mask_0, x = mixed_29_cast_fp16)[name = string("op_3851_cast_fp16")];
            int32 var_3853 = const()[name = string("op_3853"), val = int32(1)];
            bool mixed_31_interleave_0 = const()[name = string("mixed_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_31_cast_fp16 = concat(axis = var_3853, interleave = mixed_31_interleave_0, values = (var_3836_cast_fp16, var_3851_cast_fp16))[name = string("mixed_31_cast_fp16")];
            tensor<int32, [4]> var_3857_perm_0 = const()[name = string("op_3857_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_3861 = const()[name = string("op_3861"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_3857_cast_fp16 = transpose(perm = var_3857_perm_0, x = mixed_31_cast_fp16)[name = string("transpose_100")];
            tensor<fp16, [1, 16, 2048]> input_73_cast_fp16 = reshape(shape = var_3861, x = var_3857_cast_fp16)[name = string("input_73_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_7_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(425767104))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427864320))))[name = string("model_layers_7_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_52_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_self_attn_o_proj_weight_to_fp16_quantized, x = input_73_cast_fp16)[name = string("linear_52_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_95_cast_fp16 = add(x = x_85_cast_fp16, y = linear_52_cast_fp16)[name = string("x_95_cast_fp16")];
            fp16 var_3872_promoted_to_fp16 = const()[name = string("op_3872_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3874_cast_fp16 = pow(x = x_95_cast_fp16, y = var_3872_promoted_to_fp16)[name = string("op_3874_cast_fp16")];
            tensor<int32, [1]> var_3876_axes_0 = const()[name = string("op_3876_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3876_keep_dims_0 = const()[name = string("op_3876_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3876_cast_fp16 = reduce_mean(axes = var_3876_axes_0, keep_dims = var_3876_keep_dims_0, x = var_3874_cast_fp16)[name = string("op_3876_cast_fp16")];
            fp16 var_3877_to_fp16 = const()[name = string("op_3877_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3878_cast_fp16 = add(x = var_3876_cast_fp16, y = var_3877_to_fp16)[name = string("op_3878_cast_fp16")];
            fp32 var_3879_epsilon_0 = const()[name = string("op_3879_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3879_cast_fp16 = rsqrt(epsilon = var_3879_epsilon_0, x = var_3878_cast_fp16)[name = string("op_3879_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3880_cast_fp16 = mul(x = x_95_cast_fp16, y = var_3879_cast_fp16)[name = string("op_3880_cast_fp16")];
            tensor<fp16, [1024]> model_layers_7_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_7_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427866432)))];
            tensor<fp16, [1, 16, 1024]> input_75_cast_fp16 = mul(x = var_3880_cast_fp16, y = model_layers_7_post_attention_layernorm_weight_to_fp16)[name = string("input_75_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(427868544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431014336))))[name = string("model_layers_7_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_53_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_gate_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_53_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_3885_cast_fp16 = silu(x = linear_53_cast_fp16)[name = string("op_3885_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_7_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(431020544))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434166336))))[name = string("model_layers_7_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_54_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_7_mlp_up_proj_weight_to_fp16_quantized, x = input_75_cast_fp16)[name = string("linear_54_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_79_cast_fp16 = mul(x = var_3885_cast_fp16, y = linear_54_cast_fp16)[name = string("input_79_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_7_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(434172544))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437318336))))[name = string("model_layers_7_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_55_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_7_mlp_down_proj_weight_to_fp16_quantized, x = input_79_cast_fp16)[name = string("linear_55_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_97_cast_fp16 = add(x = x_95_cast_fp16, y = linear_55_cast_fp16)[name = string("x_97_cast_fp16")];
            fp16 var_3899_promoted_to_fp16 = const()[name = string("op_3899_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_3901_cast_fp16 = pow(x = x_97_cast_fp16, y = var_3899_promoted_to_fp16)[name = string("op_3901_cast_fp16")];
            tensor<int32, [1]> var_3903_axes_0 = const()[name = string("op_3903_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3903_keep_dims_0 = const()[name = string("op_3903_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_3903_cast_fp16 = reduce_mean(axes = var_3903_axes_0, keep_dims = var_3903_keep_dims_0, x = var_3901_cast_fp16)[name = string("op_3903_cast_fp16")];
            fp16 var_3904_to_fp16 = const()[name = string("op_3904_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_3905_cast_fp16 = add(x = var_3903_cast_fp16, y = var_3904_to_fp16)[name = string("op_3905_cast_fp16")];
            fp32 var_3906_epsilon_0 = const()[name = string("op_3906_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_3906_cast_fp16 = rsqrt(epsilon = var_3906_epsilon_0, x = var_3905_cast_fp16)[name = string("op_3906_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_3907_cast_fp16 = mul(x = x_97_cast_fp16, y = var_3906_cast_fp16)[name = string("op_3907_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437320448)))];
            tensor<fp16, [1, 16, 1024]> input_81_cast_fp16 = mul(x = var_3907_cast_fp16, y = model_layers_8_input_layernorm_weight_to_fp16)[name = string("input_81_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_8_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(437322560))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439419776))))[name = string("model_layers_8_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_56_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_8_self_attn_q_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_56_cast_fp16")];
            tensor<int32, [4]> var_3916 = const()[name = string("op_3916"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_3917_cast_fp16 = reshape(shape = var_3916, x = linear_56_cast_fp16)[name = string("op_3917_cast_fp16")];
            tensor<int32, [4]> x_99_perm_0 = const()[name = string("x_99_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3925_promoted_to_fp16 = const()[name = string("op_3925_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_99_cast_fp16 = transpose(perm = x_99_perm_0, x = var_3917_cast_fp16)[name = string("transpose_99")];
            tensor<fp16, [1, 16, 16, 128]> var_3927_cast_fp16 = pow(x = x_99_cast_fp16, y = var_3925_promoted_to_fp16)[name = string("op_3927_cast_fp16")];
            tensor<int32, [1]> var_3929_axes_0 = const()[name = string("op_3929_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_3929_keep_dims_0 = const()[name = string("op_3929_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_3929_cast_fp16 = reduce_mean(axes = var_3929_axes_0, keep_dims = var_3929_keep_dims_0, x = var_3927_cast_fp16)[name = string("op_3929_cast_fp16")];
            fp16 var_3930_to_fp16 = const()[name = string("op_3930_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_3931_cast_fp16 = add(x = var_3929_cast_fp16, y = var_3930_to_fp16)[name = string("op_3931_cast_fp16")];
            fp32 var_3932_epsilon_0 = const()[name = string("op_3932_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_3932_cast_fp16 = rsqrt(epsilon = var_3932_epsilon_0, x = var_3931_cast_fp16)[name = string("op_3932_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_3933_cast_fp16 = mul(x = x_99_cast_fp16, y = var_3932_cast_fp16)[name = string("op_3933_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439423936)))];
            tensor<fp16, [1, 16, 16, 128]> x_103_cast_fp16 = mul(x = var_3933_cast_fp16, y = model_layers_8_self_attn_q_norm_weight_to_fp16)[name = string("x_103_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(439424256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440472896))))[name = string("model_layers_8_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_57_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_k_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_57_cast_fp16")];
            tensor<int32, [4]> var_3942 = const()[name = string("op_3942"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3943_cast_fp16 = reshape(shape = var_3942, x = linear_57_cast_fp16)[name = string("op_3943_cast_fp16")];
            tensor<int32, [4]> x_101_perm_0 = const()[name = string("x_101_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_3951_promoted_to_fp16 = const()[name = string("op_3951_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_101_cast_fp16 = transpose(perm = x_101_perm_0, x = var_3943_cast_fp16)[name = string("transpose_98")];
            tensor<fp16, [1, 8, 16, 128]> var_3953_cast_fp16 = pow(x = x_101_cast_fp16, y = var_3951_promoted_to_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]>([-1])];
            bool var_3955_keep_dims_0 = const()[name = string("op_3955_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_3955_cast_fp16 = reduce_mean(axes = var_3955_axes_0, keep_dims = var_3955_keep_dims_0, x = var_3953_cast_fp16)[name = string("op_3955_cast_fp16")];
            fp16 var_3956_to_fp16 = const()[name = string("op_3956_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_3957_cast_fp16 = add(x = var_3955_cast_fp16, y = var_3956_to_fp16)[name = string("op_3957_cast_fp16")];
            fp32 var_3958_epsilon_0 = const()[name = string("op_3958_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_3958_cast_fp16 = rsqrt(epsilon = var_3958_epsilon_0, x = var_3957_cast_fp16)[name = string("op_3958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_3959_cast_fp16 = mul(x = x_101_cast_fp16, y = var_3958_cast_fp16)[name = string("op_3959_cast_fp16")];
            tensor<fp16, [128]> model_layers_8_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_8_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475008)))];
            tensor<fp16, [1, 8, 16, 128]> x_105_cast_fp16 = mul(x = var_3959_cast_fp16, y = model_layers_8_self_attn_k_norm_weight_to_fp16)[name = string("x_105_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_8_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(440475328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441523968))))[name = string("model_layers_8_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_58_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_v_proj_weight_to_fp16_quantized, x = input_81_cast_fp16)[name = string("linear_58_cast_fp16")];
            tensor<int32, [4]> var_3968 = const()[name = string("op_3968"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_3969_cast_fp16 = reshape(shape = var_3968, x = linear_58_cast_fp16)[name = string("op_3969_cast_fp16")];
            tensor<int32, [4]> v_17_perm_0 = const()[name = string("v_17_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_33_begin_0 = const()[name = string("first_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_33_end_0 = const()[name = string("first_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_33_end_mask_0 = const()[name = string("first_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_33_cast_fp16 = slice_by_index(begin = first_33_begin_0, end = first_33_end_0, end_mask = first_33_end_mask_0, x = x_103_cast_fp16)[name = string("first_33_cast_fp16")];
            tensor<int32, [4]> second_33_begin_0 = const()[name = string("second_33_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_33_end_0 = const()[name = string("second_33_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_33_end_mask_0 = const()[name = string("second_33_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_33_cast_fp16 = slice_by_index(begin = second_33_begin_0, end = second_33_end_0, end_mask = second_33_end_mask_0, x = x_103_cast_fp16)[name = string("second_33_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3983_cast_fp16 = mul(x = first_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3983_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3984_cast_fp16 = mul(x = second_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3984_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3986_cast_fp16 = sub(x = var_3983_cast_fp16, y = var_3984_cast_fp16)[name = string("op_3986_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3987_cast_fp16 = mul(x = second_33_cast_fp16, y = cos_1_cast_fp16)[name = string("op_3987_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3988_cast_fp16 = mul(x = first_33_cast_fp16, y = sin_1_cast_fp16)[name = string("op_3988_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_3990_cast_fp16 = add(x = var_3987_cast_fp16, y = var_3988_cast_fp16)[name = string("op_3990_cast_fp16")];
            int32 var_3992 = const()[name = string("op_3992"), val = int32(-1)];
            bool q_17_interleave_0 = const()[name = string("q_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_17_cast_fp16 = concat(axis = var_3992, interleave = q_17_interleave_0, values = (var_3986_cast_fp16, var_3990_cast_fp16))[name = string("q_17_cast_fp16")];
            tensor<int32, [4]> first_35_begin_0 = const()[name = string("first_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_35_end_0 = const()[name = string("first_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_35_end_mask_0 = const()[name = string("first_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_35_cast_fp16 = slice_by_index(begin = first_35_begin_0, end = first_35_end_0, end_mask = first_35_end_mask_0, x = x_105_cast_fp16)[name = string("first_35_cast_fp16")];
            tensor<int32, [4]> second_35_begin_0 = const()[name = string("second_35_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_35_end_0 = const()[name = string("second_35_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_35_end_mask_0 = const()[name = string("second_35_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_35_cast_fp16 = slice_by_index(begin = second_35_begin_0, end = second_35_end_0, end_mask = second_35_end_mask_0, x = x_105_cast_fp16)[name = string("second_35_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4004_cast_fp16 = mul(x = first_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4004_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4005_cast_fp16 = mul(x = second_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4005_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4007_cast_fp16 = sub(x = var_4004_cast_fp16, y = var_4005_cast_fp16)[name = string("op_4007_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4008_cast_fp16 = mul(x = second_35_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4008_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4009_cast_fp16 = mul(x = first_35_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4009_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4011_cast_fp16 = add(x = var_4008_cast_fp16, y = var_4009_cast_fp16)[name = string("op_4011_cast_fp16")];
            int32 var_4013 = const()[name = string("op_4013"), val = int32(-1)];
            bool k_17_interleave_0 = const()[name = string("k_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_17_cast_fp16 = concat(axis = var_4013, interleave = k_17_interleave_0, values = (var_4007_cast_fp16, var_4011_cast_fp16))[name = string("k_17_cast_fp16")];
            tensor<int32, [4]> var_4024_begin_0 = const()[name = string("op_4024_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4024_end_0 = const()[name = string("op_4024_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4024_end_mask_0 = const()[name = string("op_4024_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4024_stride_0 = const()[name = string("op_4024_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4024_cast_fp16 = slice_by_index(begin = var_4024_begin_0, end = var_4024_end_0, end_mask = var_4024_end_mask_0, stride = var_4024_stride_0, x = q_17_cast_fp16)[name = string("op_4024_cast_fp16")];
            tensor<int32, [4]> var_4034_begin_0 = const()[name = string("op_4034_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4034_end_0 = const()[name = string("op_4034_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4034_end_mask_0 = const()[name = string("op_4034_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4034_stride_0 = const()[name = string("op_4034_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4034_cast_fp16 = slice_by_index(begin = var_4034_begin_0, end = var_4034_end_0, end_mask = var_4034_end_mask_0, stride = var_4034_stride_0, x = q_17_cast_fp16)[name = string("op_4034_cast_fp16")];
            int32 var_4036 = const()[name = string("op_4036"), val = int32(2)];
            bool stacked_17_interleave_0 = const()[name = string("stacked_17_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_17_cast_fp16 = concat(axis = var_4036, interleave = stacked_17_interleave_0, values = (var_4024_cast_fp16, var_4034_cast_fp16))[name = string("stacked_17_cast_fp16")];
            tensor<int32, [4]> var_4042_begin_0 = const()[name = string("op_4042_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4042_end_0 = const()[name = string("op_4042_end_0"), val = tensor<int32, [4]>([9, 8, 528, 128])];
            tensor<bool, [4]> var_4042_end_mask_0 = const()[name = string("op_4042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4042_cast_fp16 = slice_by_index(begin = var_4042_begin_0, end = var_4042_end_0, end_mask = var_4042_end_mask_0, x = read_state_0)[name = string("op_4042_cast_fp16")];
            tensor<int32, [4]> var_4052_begin_0 = const()[name = string("op_4052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4052_end_0 = const()[name = string("op_4052_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4052_end_mask_0 = const()[name = string("op_4052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4052_cast_fp16 = slice_by_index(begin = var_4052_begin_0, end = var_4052_end_0, end_mask = var_4052_end_mask_0, x = var_4042_cast_fp16)[name = string("op_4052_cast_fp16")];
            bool var_4061_transpose_x_1 = const()[name = string("op_4061_transpose_x_1"), val = bool(false)];
            bool var_4061_transpose_y_1 = const()[name = string("op_4061_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4061_cast_fp16 = matmul(transpose_x = var_4061_transpose_x_1, transpose_y = var_4061_transpose_y_1, x = var_4052_cast_fp16, y = stacked_17_cast_fp16)[name = string("op_4061_cast_fp16")];
            tensor<int32, [4]> var_4064_perm_0 = const()[name = string("op_4064_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4065_to_fp16 = const()[name = string("op_4065_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4064_cast_fp16 = transpose(perm = var_4064_perm_0, x = var_4061_cast_fp16)[name = string("transpose_97")];
            tensor<fp16, [1, 8, 32, 512]> var_4066_cast_fp16 = mul(x = var_4064_cast_fp16, y = var_4065_to_fp16)[name = string("op_4066_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_17_cast_fp16 = add(x = var_4066_cast_fp16, y = cacheMask)[name = string("cache_weights_17_cast_fp16")];
            bool var_4072_transpose_x_1 = const()[name = string("op_4072_transpose_x_1"), val = bool(false)];
            bool var_4072_transpose_y_1 = const()[name = string("op_4072_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4072_cast_fp16 = matmul(transpose_x = var_4072_transpose_x_1, transpose_y = var_4072_transpose_y_1, x = stacked_17_cast_fp16, y = k_17_cast_fp16)[name = string("op_4072_cast_fp16")];
            fp16 var_4073_to_fp16 = const()[name = string("op_4073_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4074_cast_fp16 = mul(x = var_4072_cast_fp16, y = var_4073_to_fp16)[name = string("op_4074_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_17_cast_fp16 = add(x = var_4074_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_17_cast_fp16")];
            int32 var_4078 = const()[name = string("op_4078"), val = int32(-1)];
            bool var_4079_interleave_0 = const()[name = string("op_4079_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4079_cast_fp16 = concat(axis = var_4078, interleave = var_4079_interleave_0, values = (cache_weights_17_cast_fp16, call_weights_17_cast_fp16))[name = string("op_4079_cast_fp16")];
            int32 var_4080 = const()[name = string("op_4080"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_17_cast_fp16 = softmax(axis = var_4080, x = var_4079_cast_fp16)[name = string("shares_17_cast_fp16")];
            tensor<int32, [4]> var_4087_begin_0 = const()[name = string("op_4087_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4087_end_0 = const()[name = string("op_4087_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4087_end_mask_0 = const()[name = string("op_4087_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4087_cast_fp16 = slice_by_index(begin = var_4087_begin_0, end = var_4087_end_0, end_mask = var_4087_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4087_cast_fp16")];
            tensor<int32, [4]> var_4092_begin_0 = const()[name = string("op_4092_begin_0"), val = tensor<int32, [4]>([8, 0, 0, 0])];
            tensor<int32, [4]> var_4092_end_0 = const()[name = string("op_4092_end_0"), val = tensor<int32, [4]>([9, 8, 528, 128])];
            tensor<bool, [4]> var_4092_end_mask_0 = const()[name = string("op_4092_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4092_cast_fp16 = slice_by_index(begin = var_4092_begin_0, end = var_4092_end_0, end_mask = var_4092_end_mask_0, x = read_state_1)[name = string("op_4092_cast_fp16")];
            tensor<int32, [4]> var_4102_begin_0 = const()[name = string("op_4102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4102_end_0 = const()[name = string("op_4102_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4102_end_mask_0 = const()[name = string("op_4102_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4102_cast_fp16 = slice_by_index(begin = var_4102_begin_0, end = var_4102_end_0, end_mask = var_4102_end_mask_0, x = var_4092_cast_fp16)[name = string("op_4102_cast_fp16")];
            bool var_4108_transpose_x_0 = const()[name = string("op_4108_transpose_x_0"), val = bool(false)];
            bool var_4108_transpose_y_0 = const()[name = string("op_4108_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4108_cast_fp16 = matmul(transpose_x = var_4108_transpose_x_0, transpose_y = var_4108_transpose_y_0, x = var_4087_cast_fp16, y = var_4102_cast_fp16)[name = string("op_4108_cast_fp16")];
            tensor<int32, [4]> var_4113_begin_0 = const()[name = string("op_4113_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4113_end_0 = const()[name = string("op_4113_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4113_end_mask_0 = const()[name = string("op_4113_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4113_cast_fp16 = slice_by_index(begin = var_4113_begin_0, end = var_4113_end_0, end_mask = var_4113_end_mask_0, x = shares_17_cast_fp16)[name = string("op_4113_cast_fp16")];
            bool var_4114_transpose_x_0 = const()[name = string("op_4114_transpose_x_0"), val = bool(false)];
            bool var_4114_transpose_y_0 = const()[name = string("op_4114_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_17_cast_fp16 = transpose(perm = v_17_perm_0, x = var_3969_cast_fp16)[name = string("transpose_96")];
            tensor<fp16, [1, 8, 32, 128]> var_4114_cast_fp16 = matmul(transpose_x = var_4114_transpose_x_0, transpose_y = var_4114_transpose_y_0, x = var_4113_cast_fp16, y = v_17_cast_fp16)[name = string("op_4114_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_33_cast_fp16 = add(x = var_4108_cast_fp16, y = var_4114_cast_fp16)[name = string("mixed_33_cast_fp16")];
            tensor<int32, [4]> var_4131_begin_0 = const()[name = string("op_4131_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4131_end_0 = const()[name = string("op_4131_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4131_end_mask_0 = const()[name = string("op_4131_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4131_cast_fp16 = slice_by_index(begin = var_4131_begin_0, end = var_4131_end_0, end_mask = var_4131_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4131_cast_fp16")];
            tensor<int32, [4]> var_4146_begin_0 = const()[name = string("op_4146_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4146_end_0 = const()[name = string("op_4146_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4146_end_mask_0 = const()[name = string("op_4146_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4146_cast_fp16 = slice_by_index(begin = var_4146_begin_0, end = var_4146_end_0, end_mask = var_4146_end_mask_0, x = mixed_33_cast_fp16)[name = string("op_4146_cast_fp16")];
            int32 var_4148 = const()[name = string("op_4148"), val = int32(1)];
            bool mixed_35_interleave_0 = const()[name = string("mixed_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_35_cast_fp16 = concat(axis = var_4148, interleave = mixed_35_interleave_0, values = (var_4131_cast_fp16, var_4146_cast_fp16))[name = string("mixed_35_cast_fp16")];
            tensor<int32, [4]> var_4152_perm_0 = const()[name = string("op_4152_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4156 = const()[name = string("op_4156"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4152_cast_fp16 = transpose(perm = var_4152_perm_0, x = mixed_35_cast_fp16)[name = string("transpose_95")];
            tensor<fp16, [1, 16, 2048]> input_83_cast_fp16 = reshape(shape = var_4156, x = var_4152_cast_fp16)[name = string("input_83_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_8_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(441526080))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443623296))))[name = string("model_layers_8_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_59_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_self_attn_o_proj_weight_to_fp16_quantized, x = input_83_cast_fp16)[name = string("linear_59_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_107_cast_fp16 = add(x = x_97_cast_fp16, y = linear_59_cast_fp16)[name = string("x_107_cast_fp16")];
            fp16 var_4167_promoted_to_fp16 = const()[name = string("op_4167_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4169_cast_fp16 = pow(x = x_107_cast_fp16, y = var_4167_promoted_to_fp16)[name = string("op_4169_cast_fp16")];
            tensor<int32, [1]> var_4171_axes_0 = const()[name = string("op_4171_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4171_keep_dims_0 = const()[name = string("op_4171_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4171_cast_fp16 = reduce_mean(axes = var_4171_axes_0, keep_dims = var_4171_keep_dims_0, x = var_4169_cast_fp16)[name = string("op_4171_cast_fp16")];
            fp16 var_4172_to_fp16 = const()[name = string("op_4172_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4173_cast_fp16 = add(x = var_4171_cast_fp16, y = var_4172_to_fp16)[name = string("op_4173_cast_fp16")];
            fp32 var_4174_epsilon_0 = const()[name = string("op_4174_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4174_cast_fp16 = rsqrt(epsilon = var_4174_epsilon_0, x = var_4173_cast_fp16)[name = string("op_4174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4175_cast_fp16 = mul(x = x_107_cast_fp16, y = var_4174_cast_fp16)[name = string("op_4175_cast_fp16")];
            tensor<fp16, [1024]> model_layers_8_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_8_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443625408)))];
            tensor<fp16, [1, 16, 1024]> input_85_cast_fp16 = mul(x = var_4175_cast_fp16, y = model_layers_8_post_attention_layernorm_weight_to_fp16)[name = string("input_85_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(443627520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446773312))))[name = string("model_layers_8_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_60_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_gate_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_60_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4180_cast_fp16 = silu(x = linear_60_cast_fp16)[name = string("op_4180_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_8_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(446779520))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449925312))))[name = string("model_layers_8_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_61_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_8_mlp_up_proj_weight_to_fp16_quantized, x = input_85_cast_fp16)[name = string("linear_61_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_89_cast_fp16 = mul(x = var_4180_cast_fp16, y = linear_61_cast_fp16)[name = string("input_89_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_8_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(449931520))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453077312))))[name = string("model_layers_8_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_62_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_8_mlp_down_proj_weight_to_fp16_quantized, x = input_89_cast_fp16)[name = string("linear_62_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_109_cast_fp16 = add(x = x_107_cast_fp16, y = linear_62_cast_fp16)[name = string("x_109_cast_fp16")];
            fp16 var_4194_promoted_to_fp16 = const()[name = string("op_4194_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4196_cast_fp16 = pow(x = x_109_cast_fp16, y = var_4194_promoted_to_fp16)[name = string("op_4196_cast_fp16")];
            tensor<int32, [1]> var_4198_axes_0 = const()[name = string("op_4198_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4198_keep_dims_0 = const()[name = string("op_4198_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4198_cast_fp16 = reduce_mean(axes = var_4198_axes_0, keep_dims = var_4198_keep_dims_0, x = var_4196_cast_fp16)[name = string("op_4198_cast_fp16")];
            fp16 var_4199_to_fp16 = const()[name = string("op_4199_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4200_cast_fp16 = add(x = var_4198_cast_fp16, y = var_4199_to_fp16)[name = string("op_4200_cast_fp16")];
            fp32 var_4201_epsilon_0 = const()[name = string("op_4201_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4201_cast_fp16 = rsqrt(epsilon = var_4201_epsilon_0, x = var_4200_cast_fp16)[name = string("op_4201_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4202_cast_fp16 = mul(x = x_109_cast_fp16, y = var_4201_cast_fp16)[name = string("op_4202_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453079424)))];
            tensor<fp16, [1, 16, 1024]> input_91_cast_fp16 = mul(x = var_4202_cast_fp16, y = model_layers_9_input_layernorm_weight_to_fp16)[name = string("input_91_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_9_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(453081536))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455178752))))[name = string("model_layers_9_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_63_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_9_self_attn_q_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_63_cast_fp16")];
            tensor<int32, [4]> var_4211 = const()[name = string("op_4211"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4212_cast_fp16 = reshape(shape = var_4211, x = linear_63_cast_fp16)[name = string("op_4212_cast_fp16")];
            tensor<int32, [4]> x_111_perm_0 = const()[name = string("x_111_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4220_promoted_to_fp16 = const()[name = string("op_4220_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_111_cast_fp16 = transpose(perm = x_111_perm_0, x = var_4212_cast_fp16)[name = string("transpose_94")];
            tensor<fp16, [1, 16, 16, 128]> var_4222_cast_fp16 = pow(x = x_111_cast_fp16, y = var_4220_promoted_to_fp16)[name = string("op_4222_cast_fp16")];
            tensor<int32, [1]> var_4224_axes_0 = const()[name = string("op_4224_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4224_keep_dims_0 = const()[name = string("op_4224_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4224_cast_fp16 = reduce_mean(axes = var_4224_axes_0, keep_dims = var_4224_keep_dims_0, x = var_4222_cast_fp16)[name = string("op_4224_cast_fp16")];
            fp16 var_4225_to_fp16 = const()[name = string("op_4225_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4226_cast_fp16 = add(x = var_4224_cast_fp16, y = var_4225_to_fp16)[name = string("op_4226_cast_fp16")];
            fp32 var_4227_epsilon_0 = const()[name = string("op_4227_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4227_cast_fp16 = rsqrt(epsilon = var_4227_epsilon_0, x = var_4226_cast_fp16)[name = string("op_4227_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4228_cast_fp16 = mul(x = x_111_cast_fp16, y = var_4227_cast_fp16)[name = string("op_4228_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455182912)))];
            tensor<fp16, [1, 16, 16, 128]> x_115_cast_fp16 = mul(x = var_4228_cast_fp16, y = model_layers_9_self_attn_q_norm_weight_to_fp16)[name = string("x_115_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(455183232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456231872))))[name = string("model_layers_9_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_64_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_k_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_64_cast_fp16")];
            tensor<int32, [4]> var_4237 = const()[name = string("op_4237"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4238_cast_fp16 = reshape(shape = var_4237, x = linear_64_cast_fp16)[name = string("op_4238_cast_fp16")];
            tensor<int32, [4]> x_113_perm_0 = const()[name = string("x_113_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4246_promoted_to_fp16 = const()[name = string("op_4246_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_113_cast_fp16 = transpose(perm = x_113_perm_0, x = var_4238_cast_fp16)[name = string("transpose_93")];
            tensor<fp16, [1, 8, 16, 128]> var_4248_cast_fp16 = pow(x = x_113_cast_fp16, y = var_4246_promoted_to_fp16)[name = string("op_4248_cast_fp16")];
            tensor<int32, [1]> var_4250_axes_0 = const()[name = string("op_4250_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4250_keep_dims_0 = const()[name = string("op_4250_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4250_cast_fp16 = reduce_mean(axes = var_4250_axes_0, keep_dims = var_4250_keep_dims_0, x = var_4248_cast_fp16)[name = string("op_4250_cast_fp16")];
            fp16 var_4251_to_fp16 = const()[name = string("op_4251_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4252_cast_fp16 = add(x = var_4250_cast_fp16, y = var_4251_to_fp16)[name = string("op_4252_cast_fp16")];
            fp32 var_4253_epsilon_0 = const()[name = string("op_4253_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4253_cast_fp16 = rsqrt(epsilon = var_4253_epsilon_0, x = var_4252_cast_fp16)[name = string("op_4253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4254_cast_fp16 = mul(x = x_113_cast_fp16, y = var_4253_cast_fp16)[name = string("op_4254_cast_fp16")];
            tensor<fp16, [128]> model_layers_9_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_9_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456233984)))];
            tensor<fp16, [1, 8, 16, 128]> x_117_cast_fp16 = mul(x = var_4254_cast_fp16, y = model_layers_9_self_attn_k_norm_weight_to_fp16)[name = string("x_117_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_9_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(456234304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457282944))))[name = string("model_layers_9_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_65_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_v_proj_weight_to_fp16_quantized, x = input_91_cast_fp16)[name = string("linear_65_cast_fp16")];
            tensor<int32, [4]> var_4263 = const()[name = string("op_4263"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4264_cast_fp16 = reshape(shape = var_4263, x = linear_65_cast_fp16)[name = string("op_4264_cast_fp16")];
            tensor<int32, [4]> v_19_perm_0 = const()[name = string("v_19_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_37_begin_0 = const()[name = string("first_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_37_end_0 = const()[name = string("first_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_37_end_mask_0 = const()[name = string("first_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_37_cast_fp16 = slice_by_index(begin = first_37_begin_0, end = first_37_end_0, end_mask = first_37_end_mask_0, x = x_115_cast_fp16)[name = string("first_37_cast_fp16")];
            tensor<int32, [4]> second_37_begin_0 = const()[name = string("second_37_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_37_end_0 = const()[name = string("second_37_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_37_end_mask_0 = const()[name = string("second_37_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_37_cast_fp16 = slice_by_index(begin = second_37_begin_0, end = second_37_end_0, end_mask = second_37_end_mask_0, x = x_115_cast_fp16)[name = string("second_37_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4278_cast_fp16 = mul(x = first_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4278_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4279_cast_fp16 = mul(x = second_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4279_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4281_cast_fp16 = sub(x = var_4278_cast_fp16, y = var_4279_cast_fp16)[name = string("op_4281_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4282_cast_fp16 = mul(x = second_37_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4282_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4283_cast_fp16 = mul(x = first_37_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4283_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4285_cast_fp16 = add(x = var_4282_cast_fp16, y = var_4283_cast_fp16)[name = string("op_4285_cast_fp16")];
            int32 var_4287 = const()[name = string("op_4287"), val = int32(-1)];
            bool q_19_interleave_0 = const()[name = string("q_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_19_cast_fp16 = concat(axis = var_4287, interleave = q_19_interleave_0, values = (var_4281_cast_fp16, var_4285_cast_fp16))[name = string("q_19_cast_fp16")];
            tensor<int32, [4]> first_39_begin_0 = const()[name = string("first_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_39_end_0 = const()[name = string("first_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_39_end_mask_0 = const()[name = string("first_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_39_cast_fp16 = slice_by_index(begin = first_39_begin_0, end = first_39_end_0, end_mask = first_39_end_mask_0, x = x_117_cast_fp16)[name = string("first_39_cast_fp16")];
            tensor<int32, [4]> second_39_begin_0 = const()[name = string("second_39_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_39_end_0 = const()[name = string("second_39_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_39_end_mask_0 = const()[name = string("second_39_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_39_cast_fp16 = slice_by_index(begin = second_39_begin_0, end = second_39_end_0, end_mask = second_39_end_mask_0, x = x_117_cast_fp16)[name = string("second_39_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4299_cast_fp16 = mul(x = first_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4299_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4300_cast_fp16 = mul(x = second_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4300_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4302_cast_fp16 = sub(x = var_4299_cast_fp16, y = var_4300_cast_fp16)[name = string("op_4302_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4303_cast_fp16 = mul(x = second_39_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4303_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4304_cast_fp16 = mul(x = first_39_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4304_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4306_cast_fp16 = add(x = var_4303_cast_fp16, y = var_4304_cast_fp16)[name = string("op_4306_cast_fp16")];
            int32 var_4308 = const()[name = string("op_4308"), val = int32(-1)];
            bool k_19_interleave_0 = const()[name = string("k_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_19_cast_fp16 = concat(axis = var_4308, interleave = k_19_interleave_0, values = (var_4302_cast_fp16, var_4306_cast_fp16))[name = string("k_19_cast_fp16")];
            tensor<int32, [4]> var_4319_begin_0 = const()[name = string("op_4319_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4319_end_0 = const()[name = string("op_4319_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4319_end_mask_0 = const()[name = string("op_4319_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4319_stride_0 = const()[name = string("op_4319_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4319_cast_fp16 = slice_by_index(begin = var_4319_begin_0, end = var_4319_end_0, end_mask = var_4319_end_mask_0, stride = var_4319_stride_0, x = q_19_cast_fp16)[name = string("op_4319_cast_fp16")];
            tensor<int32, [4]> var_4329_begin_0 = const()[name = string("op_4329_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4329_end_0 = const()[name = string("op_4329_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4329_end_mask_0 = const()[name = string("op_4329_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4329_stride_0 = const()[name = string("op_4329_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4329_cast_fp16 = slice_by_index(begin = var_4329_begin_0, end = var_4329_end_0, end_mask = var_4329_end_mask_0, stride = var_4329_stride_0, x = q_19_cast_fp16)[name = string("op_4329_cast_fp16")];
            int32 var_4331 = const()[name = string("op_4331"), val = int32(2)];
            bool stacked_19_interleave_0 = const()[name = string("stacked_19_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_19_cast_fp16 = concat(axis = var_4331, interleave = stacked_19_interleave_0, values = (var_4319_cast_fp16, var_4329_cast_fp16))[name = string("stacked_19_cast_fp16")];
            tensor<int32, [4]> var_4337_begin_0 = const()[name = string("op_4337_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4337_end_0 = const()[name = string("op_4337_end_0"), val = tensor<int32, [4]>([10, 8, 528, 128])];
            tensor<bool, [4]> var_4337_end_mask_0 = const()[name = string("op_4337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4337_cast_fp16 = slice_by_index(begin = var_4337_begin_0, end = var_4337_end_0, end_mask = var_4337_end_mask_0, x = read_state_0)[name = string("op_4337_cast_fp16")];
            tensor<int32, [4]> var_4347_begin_0 = const()[name = string("op_4347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4347_end_0 = const()[name = string("op_4347_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4347_end_mask_0 = const()[name = string("op_4347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4347_cast_fp16 = slice_by_index(begin = var_4347_begin_0, end = var_4347_end_0, end_mask = var_4347_end_mask_0, x = var_4337_cast_fp16)[name = string("op_4347_cast_fp16")];
            bool var_4356_transpose_x_1 = const()[name = string("op_4356_transpose_x_1"), val = bool(false)];
            bool var_4356_transpose_y_1 = const()[name = string("op_4356_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4356_cast_fp16 = matmul(transpose_x = var_4356_transpose_x_1, transpose_y = var_4356_transpose_y_1, x = var_4347_cast_fp16, y = stacked_19_cast_fp16)[name = string("op_4356_cast_fp16")];
            tensor<int32, [4]> var_4359_perm_0 = const()[name = string("op_4359_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4360_to_fp16 = const()[name = string("op_4360_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4359_cast_fp16 = transpose(perm = var_4359_perm_0, x = var_4356_cast_fp16)[name = string("transpose_92")];
            tensor<fp16, [1, 8, 32, 512]> var_4361_cast_fp16 = mul(x = var_4359_cast_fp16, y = var_4360_to_fp16)[name = string("op_4361_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_19_cast_fp16 = add(x = var_4361_cast_fp16, y = cacheMask)[name = string("cache_weights_19_cast_fp16")];
            bool var_4367_transpose_x_1 = const()[name = string("op_4367_transpose_x_1"), val = bool(false)];
            bool var_4367_transpose_y_1 = const()[name = string("op_4367_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4367_cast_fp16 = matmul(transpose_x = var_4367_transpose_x_1, transpose_y = var_4367_transpose_y_1, x = stacked_19_cast_fp16, y = k_19_cast_fp16)[name = string("op_4367_cast_fp16")];
            fp16 var_4368_to_fp16 = const()[name = string("op_4368_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4369_cast_fp16 = mul(x = var_4367_cast_fp16, y = var_4368_to_fp16)[name = string("op_4369_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_19_cast_fp16 = add(x = var_4369_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_19_cast_fp16")];
            int32 var_4373 = const()[name = string("op_4373"), val = int32(-1)];
            bool var_4374_interleave_0 = const()[name = string("op_4374_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4374_cast_fp16 = concat(axis = var_4373, interleave = var_4374_interleave_0, values = (cache_weights_19_cast_fp16, call_weights_19_cast_fp16))[name = string("op_4374_cast_fp16")];
            int32 var_4375 = const()[name = string("op_4375"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_19_cast_fp16 = softmax(axis = var_4375, x = var_4374_cast_fp16)[name = string("shares_19_cast_fp16")];
            tensor<int32, [4]> var_4382_begin_0 = const()[name = string("op_4382_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4382_end_0 = const()[name = string("op_4382_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4382_end_mask_0 = const()[name = string("op_4382_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4382_cast_fp16 = slice_by_index(begin = var_4382_begin_0, end = var_4382_end_0, end_mask = var_4382_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4382_cast_fp16")];
            tensor<int32, [4]> var_4387_begin_0 = const()[name = string("op_4387_begin_0"), val = tensor<int32, [4]>([9, 0, 0, 0])];
            tensor<int32, [4]> var_4387_end_0 = const()[name = string("op_4387_end_0"), val = tensor<int32, [4]>([10, 8, 528, 128])];
            tensor<bool, [4]> var_4387_end_mask_0 = const()[name = string("op_4387_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4387_cast_fp16 = slice_by_index(begin = var_4387_begin_0, end = var_4387_end_0, end_mask = var_4387_end_mask_0, x = read_state_1)[name = string("op_4387_cast_fp16")];
            tensor<int32, [4]> var_4397_begin_0 = const()[name = string("op_4397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4397_end_0 = const()[name = string("op_4397_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4397_end_mask_0 = const()[name = string("op_4397_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4397_cast_fp16 = slice_by_index(begin = var_4397_begin_0, end = var_4397_end_0, end_mask = var_4397_end_mask_0, x = var_4387_cast_fp16)[name = string("op_4397_cast_fp16")];
            bool var_4403_transpose_x_0 = const()[name = string("op_4403_transpose_x_0"), val = bool(false)];
            bool var_4403_transpose_y_0 = const()[name = string("op_4403_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4403_cast_fp16 = matmul(transpose_x = var_4403_transpose_x_0, transpose_y = var_4403_transpose_y_0, x = var_4382_cast_fp16, y = var_4397_cast_fp16)[name = string("op_4403_cast_fp16")];
            tensor<int32, [4]> var_4408_begin_0 = const()[name = string("op_4408_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4408_end_0 = const()[name = string("op_4408_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4408_end_mask_0 = const()[name = string("op_4408_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4408_cast_fp16 = slice_by_index(begin = var_4408_begin_0, end = var_4408_end_0, end_mask = var_4408_end_mask_0, x = shares_19_cast_fp16)[name = string("op_4408_cast_fp16")];
            bool var_4409_transpose_x_0 = const()[name = string("op_4409_transpose_x_0"), val = bool(false)];
            bool var_4409_transpose_y_0 = const()[name = string("op_4409_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_19_cast_fp16 = transpose(perm = v_19_perm_0, x = var_4264_cast_fp16)[name = string("transpose_91")];
            tensor<fp16, [1, 8, 32, 128]> var_4409_cast_fp16 = matmul(transpose_x = var_4409_transpose_x_0, transpose_y = var_4409_transpose_y_0, x = var_4408_cast_fp16, y = v_19_cast_fp16)[name = string("op_4409_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_37_cast_fp16 = add(x = var_4403_cast_fp16, y = var_4409_cast_fp16)[name = string("mixed_37_cast_fp16")];
            tensor<int32, [4]> var_4426_begin_0 = const()[name = string("op_4426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4426_end_0 = const()[name = string("op_4426_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4426_end_mask_0 = const()[name = string("op_4426_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4426_cast_fp16 = slice_by_index(begin = var_4426_begin_0, end = var_4426_end_0, end_mask = var_4426_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4426_cast_fp16")];
            tensor<int32, [4]> var_4441_begin_0 = const()[name = string("op_4441_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4441_end_0 = const()[name = string("op_4441_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4441_end_mask_0 = const()[name = string("op_4441_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4441_cast_fp16 = slice_by_index(begin = var_4441_begin_0, end = var_4441_end_0, end_mask = var_4441_end_mask_0, x = mixed_37_cast_fp16)[name = string("op_4441_cast_fp16")];
            int32 var_4443 = const()[name = string("op_4443"), val = int32(1)];
            bool mixed_39_interleave_0 = const()[name = string("mixed_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_39_cast_fp16 = concat(axis = var_4443, interleave = mixed_39_interleave_0, values = (var_4426_cast_fp16, var_4441_cast_fp16))[name = string("mixed_39_cast_fp16")];
            tensor<int32, [4]> var_4447_perm_0 = const()[name = string("op_4447_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4451 = const()[name = string("op_4451"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4447_cast_fp16 = transpose(perm = var_4447_perm_0, x = mixed_39_cast_fp16)[name = string("transpose_90")];
            tensor<fp16, [1, 16, 2048]> input_93_cast_fp16 = reshape(shape = var_4451, x = var_4447_cast_fp16)[name = string("input_93_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_9_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457285056))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459382272))))[name = string("model_layers_9_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_66_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_self_attn_o_proj_weight_to_fp16_quantized, x = input_93_cast_fp16)[name = string("linear_66_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_119_cast_fp16 = add(x = x_109_cast_fp16, y = linear_66_cast_fp16)[name = string("x_119_cast_fp16")];
            fp16 var_4462_promoted_to_fp16 = const()[name = string("op_4462_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4464_cast_fp16 = pow(x = x_119_cast_fp16, y = var_4462_promoted_to_fp16)[name = string("op_4464_cast_fp16")];
            tensor<int32, [1]> var_4466_axes_0 = const()[name = string("op_4466_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4466_keep_dims_0 = const()[name = string("op_4466_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4466_cast_fp16 = reduce_mean(axes = var_4466_axes_0, keep_dims = var_4466_keep_dims_0, x = var_4464_cast_fp16)[name = string("op_4466_cast_fp16")];
            fp16 var_4467_to_fp16 = const()[name = string("op_4467_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4468_cast_fp16 = add(x = var_4466_cast_fp16, y = var_4467_to_fp16)[name = string("op_4468_cast_fp16")];
            fp32 var_4469_epsilon_0 = const()[name = string("op_4469_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4469_cast_fp16 = rsqrt(epsilon = var_4469_epsilon_0, x = var_4468_cast_fp16)[name = string("op_4469_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4470_cast_fp16 = mul(x = x_119_cast_fp16, y = var_4469_cast_fp16)[name = string("op_4470_cast_fp16")];
            tensor<fp16, [1024]> model_layers_9_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_9_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459384384)))];
            tensor<fp16, [1, 16, 1024]> input_95_cast_fp16 = mul(x = var_4470_cast_fp16, y = model_layers_9_post_attention_layernorm_weight_to_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(459386496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462532288))))[name = string("model_layers_9_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_67_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_gate_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_67_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4475_cast_fp16 = silu(x = linear_67_cast_fp16)[name = string("op_4475_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_9_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(462538496))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465684288))))[name = string("model_layers_9_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_68_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_9_mlp_up_proj_weight_to_fp16_quantized, x = input_95_cast_fp16)[name = string("linear_68_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_99_cast_fp16 = mul(x = var_4475_cast_fp16, y = linear_68_cast_fp16)[name = string("input_99_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_9_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(465690496))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468836288))))[name = string("model_layers_9_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_69_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_9_mlp_down_proj_weight_to_fp16_quantized, x = input_99_cast_fp16)[name = string("linear_69_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_121_cast_fp16 = add(x = x_119_cast_fp16, y = linear_69_cast_fp16)[name = string("x_121_cast_fp16")];
            fp16 var_4489_promoted_to_fp16 = const()[name = string("op_4489_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4491_cast_fp16 = pow(x = x_121_cast_fp16, y = var_4489_promoted_to_fp16)[name = string("op_4491_cast_fp16")];
            tensor<int32, [1]> var_4493_axes_0 = const()[name = string("op_4493_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4493_keep_dims_0 = const()[name = string("op_4493_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4493_cast_fp16 = reduce_mean(axes = var_4493_axes_0, keep_dims = var_4493_keep_dims_0, x = var_4491_cast_fp16)[name = string("op_4493_cast_fp16")];
            fp16 var_4494_to_fp16 = const()[name = string("op_4494_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4495_cast_fp16 = add(x = var_4493_cast_fp16, y = var_4494_to_fp16)[name = string("op_4495_cast_fp16")];
            fp32 var_4496_epsilon_0 = const()[name = string("op_4496_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4496_cast_fp16 = rsqrt(epsilon = var_4496_epsilon_0, x = var_4495_cast_fp16)[name = string("op_4496_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4497_cast_fp16 = mul(x = x_121_cast_fp16, y = var_4496_cast_fp16)[name = string("op_4497_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468838400)))];
            tensor<fp16, [1, 16, 1024]> input_101_cast_fp16 = mul(x = var_4497_cast_fp16, y = model_layers_10_input_layernorm_weight_to_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_10_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(468840512))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470937728))))[name = string("model_layers_10_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_70_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_10_self_attn_q_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_70_cast_fp16")];
            tensor<int32, [4]> var_4506 = const()[name = string("op_4506"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4507_cast_fp16 = reshape(shape = var_4506, x = linear_70_cast_fp16)[name = string("op_4507_cast_fp16")];
            tensor<int32, [4]> x_123_perm_0 = const()[name = string("x_123_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4515_promoted_to_fp16 = const()[name = string("op_4515_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_123_cast_fp16 = transpose(perm = x_123_perm_0, x = var_4507_cast_fp16)[name = string("transpose_89")];
            tensor<fp16, [1, 16, 16, 128]> var_4517_cast_fp16 = pow(x = x_123_cast_fp16, y = var_4515_promoted_to_fp16)[name = string("op_4517_cast_fp16")];
            tensor<int32, [1]> var_4519_axes_0 = const()[name = string("op_4519_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4519_keep_dims_0 = const()[name = string("op_4519_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4519_cast_fp16 = reduce_mean(axes = var_4519_axes_0, keep_dims = var_4519_keep_dims_0, x = var_4517_cast_fp16)[name = string("op_4519_cast_fp16")];
            fp16 var_4520_to_fp16 = const()[name = string("op_4520_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4521_cast_fp16 = add(x = var_4519_cast_fp16, y = var_4520_to_fp16)[name = string("op_4521_cast_fp16")];
            fp32 var_4522_epsilon_0 = const()[name = string("op_4522_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4522_cast_fp16 = rsqrt(epsilon = var_4522_epsilon_0, x = var_4521_cast_fp16)[name = string("op_4522_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4523_cast_fp16 = mul(x = x_123_cast_fp16, y = var_4522_cast_fp16)[name = string("op_4523_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470941888)))];
            tensor<fp16, [1, 16, 16, 128]> x_127_cast_fp16 = mul(x = var_4523_cast_fp16, y = model_layers_10_self_attn_q_norm_weight_to_fp16)[name = string("x_127_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(470942208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471990848))))[name = string("model_layers_10_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_71_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_k_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_71_cast_fp16")];
            tensor<int32, [4]> var_4532 = const()[name = string("op_4532"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4533_cast_fp16 = reshape(shape = var_4532, x = linear_71_cast_fp16)[name = string("op_4533_cast_fp16")];
            tensor<int32, [4]> x_125_perm_0 = const()[name = string("x_125_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4541_promoted_to_fp16 = const()[name = string("op_4541_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_125_cast_fp16 = transpose(perm = x_125_perm_0, x = var_4533_cast_fp16)[name = string("transpose_88")];
            tensor<fp16, [1, 8, 16, 128]> var_4543_cast_fp16 = pow(x = x_125_cast_fp16, y = var_4541_promoted_to_fp16)[name = string("op_4543_cast_fp16")];
            tensor<int32, [1]> var_4545_axes_0 = const()[name = string("op_4545_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4545_keep_dims_0 = const()[name = string("op_4545_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4545_cast_fp16 = reduce_mean(axes = var_4545_axes_0, keep_dims = var_4545_keep_dims_0, x = var_4543_cast_fp16)[name = string("op_4545_cast_fp16")];
            fp16 var_4546_to_fp16 = const()[name = string("op_4546_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4547_cast_fp16 = add(x = var_4545_cast_fp16, y = var_4546_to_fp16)[name = string("op_4547_cast_fp16")];
            fp32 var_4548_epsilon_0 = const()[name = string("op_4548_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4548_cast_fp16 = rsqrt(epsilon = var_4548_epsilon_0, x = var_4547_cast_fp16)[name = string("op_4548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4549_cast_fp16 = mul(x = x_125_cast_fp16, y = var_4548_cast_fp16)[name = string("op_4549_cast_fp16")];
            tensor<fp16, [128]> model_layers_10_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_10_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471992960)))];
            tensor<fp16, [1, 8, 16, 128]> x_129_cast_fp16 = mul(x = var_4549_cast_fp16, y = model_layers_10_self_attn_k_norm_weight_to_fp16)[name = string("x_129_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_10_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(471993280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473041920))))[name = string("model_layers_10_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_72_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_v_proj_weight_to_fp16_quantized, x = input_101_cast_fp16)[name = string("linear_72_cast_fp16")];
            tensor<int32, [4]> var_4558 = const()[name = string("op_4558"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4559_cast_fp16 = reshape(shape = var_4558, x = linear_72_cast_fp16)[name = string("op_4559_cast_fp16")];
            tensor<int32, [4]> v_21_perm_0 = const()[name = string("v_21_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_41_begin_0 = const()[name = string("first_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_41_end_0 = const()[name = string("first_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_41_end_mask_0 = const()[name = string("first_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_41_cast_fp16 = slice_by_index(begin = first_41_begin_0, end = first_41_end_0, end_mask = first_41_end_mask_0, x = x_127_cast_fp16)[name = string("first_41_cast_fp16")];
            tensor<int32, [4]> second_41_begin_0 = const()[name = string("second_41_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_41_end_0 = const()[name = string("second_41_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_41_end_mask_0 = const()[name = string("second_41_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_41_cast_fp16 = slice_by_index(begin = second_41_begin_0, end = second_41_end_0, end_mask = second_41_end_mask_0, x = x_127_cast_fp16)[name = string("second_41_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4573_cast_fp16 = mul(x = first_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4573_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4574_cast_fp16 = mul(x = second_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4574_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4576_cast_fp16 = sub(x = var_4573_cast_fp16, y = var_4574_cast_fp16)[name = string("op_4576_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4577_cast_fp16 = mul(x = second_41_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4577_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4578_cast_fp16 = mul(x = first_41_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4578_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4580_cast_fp16 = add(x = var_4577_cast_fp16, y = var_4578_cast_fp16)[name = string("op_4580_cast_fp16")];
            int32 var_4582 = const()[name = string("op_4582"), val = int32(-1)];
            bool q_21_interleave_0 = const()[name = string("q_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_21_cast_fp16 = concat(axis = var_4582, interleave = q_21_interleave_0, values = (var_4576_cast_fp16, var_4580_cast_fp16))[name = string("q_21_cast_fp16")];
            tensor<int32, [4]> first_43_begin_0 = const()[name = string("first_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_43_end_0 = const()[name = string("first_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_43_end_mask_0 = const()[name = string("first_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_43_cast_fp16 = slice_by_index(begin = first_43_begin_0, end = first_43_end_0, end_mask = first_43_end_mask_0, x = x_129_cast_fp16)[name = string("first_43_cast_fp16")];
            tensor<int32, [4]> second_43_begin_0 = const()[name = string("second_43_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_43_end_0 = const()[name = string("second_43_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_43_end_mask_0 = const()[name = string("second_43_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_43_cast_fp16 = slice_by_index(begin = second_43_begin_0, end = second_43_end_0, end_mask = second_43_end_mask_0, x = x_129_cast_fp16)[name = string("second_43_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4594_cast_fp16 = mul(x = first_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4594_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4595_cast_fp16 = mul(x = second_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4595_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4597_cast_fp16 = sub(x = var_4594_cast_fp16, y = var_4595_cast_fp16)[name = string("op_4597_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4598_cast_fp16 = mul(x = second_43_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4598_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4599_cast_fp16 = mul(x = first_43_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4599_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4601_cast_fp16 = add(x = var_4598_cast_fp16, y = var_4599_cast_fp16)[name = string("op_4601_cast_fp16")];
            int32 var_4603 = const()[name = string("op_4603"), val = int32(-1)];
            bool k_21_interleave_0 = const()[name = string("k_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_21_cast_fp16 = concat(axis = var_4603, interleave = k_21_interleave_0, values = (var_4597_cast_fp16, var_4601_cast_fp16))[name = string("k_21_cast_fp16")];
            tensor<int32, [4]> var_4614_begin_0 = const()[name = string("op_4614_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4614_end_0 = const()[name = string("op_4614_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4614_end_mask_0 = const()[name = string("op_4614_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4614_stride_0 = const()[name = string("op_4614_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4614_cast_fp16 = slice_by_index(begin = var_4614_begin_0, end = var_4614_end_0, end_mask = var_4614_end_mask_0, stride = var_4614_stride_0, x = q_21_cast_fp16)[name = string("op_4614_cast_fp16")];
            tensor<int32, [4]> var_4624_begin_0 = const()[name = string("op_4624_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4624_end_0 = const()[name = string("op_4624_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4624_end_mask_0 = const()[name = string("op_4624_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4624_stride_0 = const()[name = string("op_4624_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4624_cast_fp16 = slice_by_index(begin = var_4624_begin_0, end = var_4624_end_0, end_mask = var_4624_end_mask_0, stride = var_4624_stride_0, x = q_21_cast_fp16)[name = string("op_4624_cast_fp16")];
            int32 var_4626 = const()[name = string("op_4626"), val = int32(2)];
            bool stacked_21_interleave_0 = const()[name = string("stacked_21_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_21_cast_fp16 = concat(axis = var_4626, interleave = stacked_21_interleave_0, values = (var_4614_cast_fp16, var_4624_cast_fp16))[name = string("stacked_21_cast_fp16")];
            tensor<int32, [4]> var_4632_begin_0 = const()[name = string("op_4632_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4632_end_0 = const()[name = string("op_4632_end_0"), val = tensor<int32, [4]>([11, 8, 528, 128])];
            tensor<bool, [4]> var_4632_end_mask_0 = const()[name = string("op_4632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4632_cast_fp16 = slice_by_index(begin = var_4632_begin_0, end = var_4632_end_0, end_mask = var_4632_end_mask_0, x = read_state_0)[name = string("op_4632_cast_fp16")];
            tensor<int32, [4]> var_4642_begin_0 = const()[name = string("op_4642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4642_end_0 = const()[name = string("op_4642_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4642_end_mask_0 = const()[name = string("op_4642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4642_cast_fp16 = slice_by_index(begin = var_4642_begin_0, end = var_4642_end_0, end_mask = var_4642_end_mask_0, x = var_4632_cast_fp16)[name = string("op_4642_cast_fp16")];
            bool var_4651_transpose_x_1 = const()[name = string("op_4651_transpose_x_1"), val = bool(false)];
            bool var_4651_transpose_y_1 = const()[name = string("op_4651_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4651_cast_fp16 = matmul(transpose_x = var_4651_transpose_x_1, transpose_y = var_4651_transpose_y_1, x = var_4642_cast_fp16, y = stacked_21_cast_fp16)[name = string("op_4651_cast_fp16")];
            tensor<int32, [4]> var_4654_perm_0 = const()[name = string("op_4654_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4655_to_fp16 = const()[name = string("op_4655_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4654_cast_fp16 = transpose(perm = var_4654_perm_0, x = var_4651_cast_fp16)[name = string("transpose_87")];
            tensor<fp16, [1, 8, 32, 512]> var_4656_cast_fp16 = mul(x = var_4654_cast_fp16, y = var_4655_to_fp16)[name = string("op_4656_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_21_cast_fp16 = add(x = var_4656_cast_fp16, y = cacheMask)[name = string("cache_weights_21_cast_fp16")];
            bool var_4662_transpose_x_1 = const()[name = string("op_4662_transpose_x_1"), val = bool(false)];
            bool var_4662_transpose_y_1 = const()[name = string("op_4662_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4662_cast_fp16 = matmul(transpose_x = var_4662_transpose_x_1, transpose_y = var_4662_transpose_y_1, x = stacked_21_cast_fp16, y = k_21_cast_fp16)[name = string("op_4662_cast_fp16")];
            fp16 var_4663_to_fp16 = const()[name = string("op_4663_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4664_cast_fp16 = mul(x = var_4662_cast_fp16, y = var_4663_to_fp16)[name = string("op_4664_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_21_cast_fp16 = add(x = var_4664_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_21_cast_fp16")];
            int32 var_4668 = const()[name = string("op_4668"), val = int32(-1)];
            bool var_4669_interleave_0 = const()[name = string("op_4669_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4669_cast_fp16 = concat(axis = var_4668, interleave = var_4669_interleave_0, values = (cache_weights_21_cast_fp16, call_weights_21_cast_fp16))[name = string("op_4669_cast_fp16")];
            int32 var_4670 = const()[name = string("op_4670"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_21_cast_fp16 = softmax(axis = var_4670, x = var_4669_cast_fp16)[name = string("shares_21_cast_fp16")];
            tensor<int32, [4]> var_4677_begin_0 = const()[name = string("op_4677_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4677_end_0 = const()[name = string("op_4677_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4677_end_mask_0 = const()[name = string("op_4677_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4677_cast_fp16 = slice_by_index(begin = var_4677_begin_0, end = var_4677_end_0, end_mask = var_4677_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4677_cast_fp16")];
            tensor<int32, [4]> var_4682_begin_0 = const()[name = string("op_4682_begin_0"), val = tensor<int32, [4]>([10, 0, 0, 0])];
            tensor<int32, [4]> var_4682_end_0 = const()[name = string("op_4682_end_0"), val = tensor<int32, [4]>([11, 8, 528, 128])];
            tensor<bool, [4]> var_4682_end_mask_0 = const()[name = string("op_4682_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4682_cast_fp16 = slice_by_index(begin = var_4682_begin_0, end = var_4682_end_0, end_mask = var_4682_end_mask_0, x = read_state_1)[name = string("op_4682_cast_fp16")];
            tensor<int32, [4]> var_4692_begin_0 = const()[name = string("op_4692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4692_end_0 = const()[name = string("op_4692_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4692_end_mask_0 = const()[name = string("op_4692_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4692_cast_fp16 = slice_by_index(begin = var_4692_begin_0, end = var_4692_end_0, end_mask = var_4692_end_mask_0, x = var_4682_cast_fp16)[name = string("op_4692_cast_fp16")];
            bool var_4698_transpose_x_0 = const()[name = string("op_4698_transpose_x_0"), val = bool(false)];
            bool var_4698_transpose_y_0 = const()[name = string("op_4698_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4698_cast_fp16 = matmul(transpose_x = var_4698_transpose_x_0, transpose_y = var_4698_transpose_y_0, x = var_4677_cast_fp16, y = var_4692_cast_fp16)[name = string("op_4698_cast_fp16")];
            tensor<int32, [4]> var_4703_begin_0 = const()[name = string("op_4703_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4703_end_0 = const()[name = string("op_4703_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4703_end_mask_0 = const()[name = string("op_4703_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4703_cast_fp16 = slice_by_index(begin = var_4703_begin_0, end = var_4703_end_0, end_mask = var_4703_end_mask_0, x = shares_21_cast_fp16)[name = string("op_4703_cast_fp16")];
            bool var_4704_transpose_x_0 = const()[name = string("op_4704_transpose_x_0"), val = bool(false)];
            bool var_4704_transpose_y_0 = const()[name = string("op_4704_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_21_cast_fp16 = transpose(perm = v_21_perm_0, x = var_4559_cast_fp16)[name = string("transpose_86")];
            tensor<fp16, [1, 8, 32, 128]> var_4704_cast_fp16 = matmul(transpose_x = var_4704_transpose_x_0, transpose_y = var_4704_transpose_y_0, x = var_4703_cast_fp16, y = v_21_cast_fp16)[name = string("op_4704_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_41_cast_fp16 = add(x = var_4698_cast_fp16, y = var_4704_cast_fp16)[name = string("mixed_41_cast_fp16")];
            tensor<int32, [4]> var_4721_begin_0 = const()[name = string("op_4721_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4721_end_0 = const()[name = string("op_4721_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_4721_end_mask_0 = const()[name = string("op_4721_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4721_cast_fp16 = slice_by_index(begin = var_4721_begin_0, end = var_4721_end_0, end_mask = var_4721_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4721_cast_fp16")];
            tensor<int32, [4]> var_4736_begin_0 = const()[name = string("op_4736_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_4736_end_0 = const()[name = string("op_4736_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_4736_end_mask_0 = const()[name = string("op_4736_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_4736_cast_fp16 = slice_by_index(begin = var_4736_begin_0, end = var_4736_end_0, end_mask = var_4736_end_mask_0, x = mixed_41_cast_fp16)[name = string("op_4736_cast_fp16")];
            int32 var_4738 = const()[name = string("op_4738"), val = int32(1)];
            bool mixed_43_interleave_0 = const()[name = string("mixed_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_43_cast_fp16 = concat(axis = var_4738, interleave = mixed_43_interleave_0, values = (var_4721_cast_fp16, var_4736_cast_fp16))[name = string("mixed_43_cast_fp16")];
            tensor<int32, [4]> var_4742_perm_0 = const()[name = string("op_4742_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_4746 = const()[name = string("op_4746"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_4742_cast_fp16 = transpose(perm = var_4742_perm_0, x = mixed_43_cast_fp16)[name = string("transpose_85")];
            tensor<fp16, [1, 16, 2048]> input_103_cast_fp16 = reshape(shape = var_4746, x = var_4742_cast_fp16)[name = string("input_103_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_10_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(473044032))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475141248))))[name = string("model_layers_10_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_73_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_self_attn_o_proj_weight_to_fp16_quantized, x = input_103_cast_fp16)[name = string("linear_73_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_131_cast_fp16 = add(x = x_121_cast_fp16, y = linear_73_cast_fp16)[name = string("x_131_cast_fp16")];
            fp16 var_4757_promoted_to_fp16 = const()[name = string("op_4757_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4759_cast_fp16 = pow(x = x_131_cast_fp16, y = var_4757_promoted_to_fp16)[name = string("op_4759_cast_fp16")];
            tensor<int32, [1]> var_4761_axes_0 = const()[name = string("op_4761_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4761_keep_dims_0 = const()[name = string("op_4761_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4761_cast_fp16 = reduce_mean(axes = var_4761_axes_0, keep_dims = var_4761_keep_dims_0, x = var_4759_cast_fp16)[name = string("op_4761_cast_fp16")];
            fp16 var_4762_to_fp16 = const()[name = string("op_4762_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4763_cast_fp16 = add(x = var_4761_cast_fp16, y = var_4762_to_fp16)[name = string("op_4763_cast_fp16")];
            fp32 var_4764_epsilon_0 = const()[name = string("op_4764_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4764_cast_fp16 = rsqrt(epsilon = var_4764_epsilon_0, x = var_4763_cast_fp16)[name = string("op_4764_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4765_cast_fp16 = mul(x = x_131_cast_fp16, y = var_4764_cast_fp16)[name = string("op_4765_cast_fp16")];
            tensor<fp16, [1024]> model_layers_10_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_10_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475143360)))];
            tensor<fp16, [1, 16, 1024]> input_105_cast_fp16 = mul(x = var_4765_cast_fp16, y = model_layers_10_post_attention_layernorm_weight_to_fp16)[name = string("input_105_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(475145472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478291264))))[name = string("model_layers_10_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_74_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_gate_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_74_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_4770_cast_fp16 = silu(x = linear_74_cast_fp16)[name = string("op_4770_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_10_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(478297472))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481443264))))[name = string("model_layers_10_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_75_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_10_mlp_up_proj_weight_to_fp16_quantized, x = input_105_cast_fp16)[name = string("linear_75_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_109_cast_fp16 = mul(x = var_4770_cast_fp16, y = linear_75_cast_fp16)[name = string("input_109_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_10_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(481449472))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484595264))))[name = string("model_layers_10_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_76_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_10_mlp_down_proj_weight_to_fp16_quantized, x = input_109_cast_fp16)[name = string("linear_76_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_133_cast_fp16 = add(x = x_131_cast_fp16, y = linear_76_cast_fp16)[name = string("x_133_cast_fp16")];
            fp16 var_4784_promoted_to_fp16 = const()[name = string("op_4784_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_4786_cast_fp16 = pow(x = x_133_cast_fp16, y = var_4784_promoted_to_fp16)[name = string("op_4786_cast_fp16")];
            tensor<int32, [1]> var_4788_axes_0 = const()[name = string("op_4788_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4788_keep_dims_0 = const()[name = string("op_4788_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_4788_cast_fp16 = reduce_mean(axes = var_4788_axes_0, keep_dims = var_4788_keep_dims_0, x = var_4786_cast_fp16)[name = string("op_4788_cast_fp16")];
            fp16 var_4789_to_fp16 = const()[name = string("op_4789_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_4790_cast_fp16 = add(x = var_4788_cast_fp16, y = var_4789_to_fp16)[name = string("op_4790_cast_fp16")];
            fp32 var_4791_epsilon_0 = const()[name = string("op_4791_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_4791_cast_fp16 = rsqrt(epsilon = var_4791_epsilon_0, x = var_4790_cast_fp16)[name = string("op_4791_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_4792_cast_fp16 = mul(x = x_133_cast_fp16, y = var_4791_cast_fp16)[name = string("op_4792_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484597376)))];
            tensor<fp16, [1, 16, 1024]> input_111_cast_fp16 = mul(x = var_4792_cast_fp16, y = model_layers_11_input_layernorm_weight_to_fp16)[name = string("input_111_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_11_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(484599488))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486696704))))[name = string("model_layers_11_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_77_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_11_self_attn_q_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_77_cast_fp16")];
            tensor<int32, [4]> var_4801 = const()[name = string("op_4801"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_4802_cast_fp16 = reshape(shape = var_4801, x = linear_77_cast_fp16)[name = string("op_4802_cast_fp16")];
            tensor<int32, [4]> x_135_perm_0 = const()[name = string("x_135_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4810_promoted_to_fp16 = const()[name = string("op_4810_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_135_cast_fp16 = transpose(perm = x_135_perm_0, x = var_4802_cast_fp16)[name = string("transpose_84")];
            tensor<fp16, [1, 16, 16, 128]> var_4812_cast_fp16 = pow(x = x_135_cast_fp16, y = var_4810_promoted_to_fp16)[name = string("op_4812_cast_fp16")];
            tensor<int32, [1]> var_4814_axes_0 = const()[name = string("op_4814_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4814_keep_dims_0 = const()[name = string("op_4814_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_4814_cast_fp16 = reduce_mean(axes = var_4814_axes_0, keep_dims = var_4814_keep_dims_0, x = var_4812_cast_fp16)[name = string("op_4814_cast_fp16")];
            fp16 var_4815_to_fp16 = const()[name = string("op_4815_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_4816_cast_fp16 = add(x = var_4814_cast_fp16, y = var_4815_to_fp16)[name = string("op_4816_cast_fp16")];
            fp32 var_4817_epsilon_0 = const()[name = string("op_4817_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_4817_cast_fp16 = rsqrt(epsilon = var_4817_epsilon_0, x = var_4816_cast_fp16)[name = string("op_4817_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_4818_cast_fp16 = mul(x = x_135_cast_fp16, y = var_4817_cast_fp16)[name = string("op_4818_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486700864)))];
            tensor<fp16, [1, 16, 16, 128]> x_139_cast_fp16 = mul(x = var_4818_cast_fp16, y = model_layers_11_self_attn_q_norm_weight_to_fp16)[name = string("x_139_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(486701184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487749824))))[name = string("model_layers_11_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_78_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_k_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_78_cast_fp16")];
            tensor<int32, [4]> var_4827 = const()[name = string("op_4827"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4828_cast_fp16 = reshape(shape = var_4827, x = linear_78_cast_fp16)[name = string("op_4828_cast_fp16")];
            tensor<int32, [4]> x_137_perm_0 = const()[name = string("x_137_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_4836_promoted_to_fp16 = const()[name = string("op_4836_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_137_cast_fp16 = transpose(perm = x_137_perm_0, x = var_4828_cast_fp16)[name = string("transpose_83")];
            tensor<fp16, [1, 8, 16, 128]> var_4838_cast_fp16 = pow(x = x_137_cast_fp16, y = var_4836_promoted_to_fp16)[name = string("op_4838_cast_fp16")];
            tensor<int32, [1]> var_4840_axes_0 = const()[name = string("op_4840_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_4840_keep_dims_0 = const()[name = string("op_4840_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_4840_cast_fp16 = reduce_mean(axes = var_4840_axes_0, keep_dims = var_4840_keep_dims_0, x = var_4838_cast_fp16)[name = string("op_4840_cast_fp16")];
            fp16 var_4841_to_fp16 = const()[name = string("op_4841_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_4842_cast_fp16 = add(x = var_4840_cast_fp16, y = var_4841_to_fp16)[name = string("op_4842_cast_fp16")];
            fp32 var_4843_epsilon_0 = const()[name = string("op_4843_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_4843_cast_fp16 = rsqrt(epsilon = var_4843_epsilon_0, x = var_4842_cast_fp16)[name = string("op_4843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_4844_cast_fp16 = mul(x = x_137_cast_fp16, y = var_4843_cast_fp16)[name = string("op_4844_cast_fp16")];
            tensor<fp16, [128]> model_layers_11_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_11_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487751936)))];
            tensor<fp16, [1, 8, 16, 128]> x_141_cast_fp16 = mul(x = var_4844_cast_fp16, y = model_layers_11_self_attn_k_norm_weight_to_fp16)[name = string("x_141_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_11_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(487752256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488800896))))[name = string("model_layers_11_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_79_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_v_proj_weight_to_fp16_quantized, x = input_111_cast_fp16)[name = string("linear_79_cast_fp16")];
            tensor<int32, [4]> var_4853 = const()[name = string("op_4853"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_4854_cast_fp16 = reshape(shape = var_4853, x = linear_79_cast_fp16)[name = string("op_4854_cast_fp16")];
            tensor<int32, [4]> v_23_perm_0 = const()[name = string("v_23_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_45_begin_0 = const()[name = string("first_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_45_end_0 = const()[name = string("first_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_45_end_mask_0 = const()[name = string("first_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_45_cast_fp16 = slice_by_index(begin = first_45_begin_0, end = first_45_end_0, end_mask = first_45_end_mask_0, x = x_139_cast_fp16)[name = string("first_45_cast_fp16")];
            tensor<int32, [4]> second_45_begin_0 = const()[name = string("second_45_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_45_end_0 = const()[name = string("second_45_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_45_end_mask_0 = const()[name = string("second_45_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_45_cast_fp16 = slice_by_index(begin = second_45_begin_0, end = second_45_end_0, end_mask = second_45_end_mask_0, x = x_139_cast_fp16)[name = string("second_45_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4868_cast_fp16 = mul(x = first_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4868_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4869_cast_fp16 = mul(x = second_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4869_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4871_cast_fp16 = sub(x = var_4868_cast_fp16, y = var_4869_cast_fp16)[name = string("op_4871_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4872_cast_fp16 = mul(x = second_45_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4872_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4873_cast_fp16 = mul(x = first_45_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4873_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_4875_cast_fp16 = add(x = var_4872_cast_fp16, y = var_4873_cast_fp16)[name = string("op_4875_cast_fp16")];
            int32 var_4877 = const()[name = string("op_4877"), val = int32(-1)];
            bool q_23_interleave_0 = const()[name = string("q_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_23_cast_fp16 = concat(axis = var_4877, interleave = q_23_interleave_0, values = (var_4871_cast_fp16, var_4875_cast_fp16))[name = string("q_23_cast_fp16")];
            tensor<int32, [4]> first_47_begin_0 = const()[name = string("first_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_47_end_0 = const()[name = string("first_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_47_end_mask_0 = const()[name = string("first_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_47_cast_fp16 = slice_by_index(begin = first_47_begin_0, end = first_47_end_0, end_mask = first_47_end_mask_0, x = x_141_cast_fp16)[name = string("first_47_cast_fp16")];
            tensor<int32, [4]> second_47_begin_0 = const()[name = string("second_47_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_47_end_0 = const()[name = string("second_47_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_47_end_mask_0 = const()[name = string("second_47_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_47_cast_fp16 = slice_by_index(begin = second_47_begin_0, end = second_47_end_0, end_mask = second_47_end_mask_0, x = x_141_cast_fp16)[name = string("second_47_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4889_cast_fp16 = mul(x = first_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4889_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4890_cast_fp16 = mul(x = second_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4890_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4892_cast_fp16 = sub(x = var_4889_cast_fp16, y = var_4890_cast_fp16)[name = string("op_4892_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4893_cast_fp16 = mul(x = second_47_cast_fp16, y = cos_1_cast_fp16)[name = string("op_4893_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4894_cast_fp16 = mul(x = first_47_cast_fp16, y = sin_1_cast_fp16)[name = string("op_4894_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_4896_cast_fp16 = add(x = var_4893_cast_fp16, y = var_4894_cast_fp16)[name = string("op_4896_cast_fp16")];
            int32 var_4898 = const()[name = string("op_4898"), val = int32(-1)];
            bool k_23_interleave_0 = const()[name = string("k_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_23_cast_fp16 = concat(axis = var_4898, interleave = k_23_interleave_0, values = (var_4892_cast_fp16, var_4896_cast_fp16))[name = string("k_23_cast_fp16")];
            tensor<int32, [4]> var_4909_begin_0 = const()[name = string("op_4909_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4909_end_0 = const()[name = string("op_4909_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4909_end_mask_0 = const()[name = string("op_4909_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4909_stride_0 = const()[name = string("op_4909_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4909_cast_fp16 = slice_by_index(begin = var_4909_begin_0, end = var_4909_end_0, end_mask = var_4909_end_mask_0, stride = var_4909_stride_0, x = q_23_cast_fp16)[name = string("op_4909_cast_fp16")];
            tensor<int32, [4]> var_4919_begin_0 = const()[name = string("op_4919_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_4919_end_0 = const()[name = string("op_4919_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_4919_end_mask_0 = const()[name = string("op_4919_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_4919_stride_0 = const()[name = string("op_4919_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_4919_cast_fp16 = slice_by_index(begin = var_4919_begin_0, end = var_4919_end_0, end_mask = var_4919_end_mask_0, stride = var_4919_stride_0, x = q_23_cast_fp16)[name = string("op_4919_cast_fp16")];
            int32 var_4921 = const()[name = string("op_4921"), val = int32(2)];
            bool stacked_23_interleave_0 = const()[name = string("stacked_23_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_23_cast_fp16 = concat(axis = var_4921, interleave = stacked_23_interleave_0, values = (var_4909_cast_fp16, var_4919_cast_fp16))[name = string("stacked_23_cast_fp16")];
            tensor<int32, [4]> var_4927_begin_0 = const()[name = string("op_4927_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4927_end_0 = const()[name = string("op_4927_end_0"), val = tensor<int32, [4]>([12, 8, 528, 128])];
            tensor<bool, [4]> var_4927_end_mask_0 = const()[name = string("op_4927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4927_cast_fp16 = slice_by_index(begin = var_4927_begin_0, end = var_4927_end_0, end_mask = var_4927_end_mask_0, x = read_state_0)[name = string("op_4927_cast_fp16")];
            tensor<int32, [4]> var_4937_begin_0 = const()[name = string("op_4937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4937_end_0 = const()[name = string("op_4937_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4937_end_mask_0 = const()[name = string("op_4937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4937_cast_fp16 = slice_by_index(begin = var_4937_begin_0, end = var_4937_end_0, end_mask = var_4937_end_mask_0, x = var_4927_cast_fp16)[name = string("op_4937_cast_fp16")];
            bool var_4946_transpose_x_1 = const()[name = string("op_4946_transpose_x_1"), val = bool(false)];
            bool var_4946_transpose_y_1 = const()[name = string("op_4946_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_4946_cast_fp16 = matmul(transpose_x = var_4946_transpose_x_1, transpose_y = var_4946_transpose_y_1, x = var_4937_cast_fp16, y = stacked_23_cast_fp16)[name = string("op_4946_cast_fp16")];
            tensor<int32, [4]> var_4949_perm_0 = const()[name = string("op_4949_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_4950_to_fp16 = const()[name = string("op_4950_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_4949_cast_fp16 = transpose(perm = var_4949_perm_0, x = var_4946_cast_fp16)[name = string("transpose_82")];
            tensor<fp16, [1, 8, 32, 512]> var_4951_cast_fp16 = mul(x = var_4949_cast_fp16, y = var_4950_to_fp16)[name = string("op_4951_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_23_cast_fp16 = add(x = var_4951_cast_fp16, y = cacheMask)[name = string("cache_weights_23_cast_fp16")];
            bool var_4957_transpose_x_1 = const()[name = string("op_4957_transpose_x_1"), val = bool(false)];
            bool var_4957_transpose_y_1 = const()[name = string("op_4957_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_4957_cast_fp16 = matmul(transpose_x = var_4957_transpose_x_1, transpose_y = var_4957_transpose_y_1, x = stacked_23_cast_fp16, y = k_23_cast_fp16)[name = string("op_4957_cast_fp16")];
            fp16 var_4958_to_fp16 = const()[name = string("op_4958_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_4959_cast_fp16 = mul(x = var_4957_cast_fp16, y = var_4958_to_fp16)[name = string("op_4959_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_23_cast_fp16 = add(x = var_4959_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_23_cast_fp16")];
            int32 var_4963 = const()[name = string("op_4963"), val = int32(-1)];
            bool var_4964_interleave_0 = const()[name = string("op_4964_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_4964_cast_fp16 = concat(axis = var_4963, interleave = var_4964_interleave_0, values = (cache_weights_23_cast_fp16, call_weights_23_cast_fp16))[name = string("op_4964_cast_fp16")];
            int32 var_4965 = const()[name = string("op_4965"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_23_cast_fp16 = softmax(axis = var_4965, x = var_4964_cast_fp16)[name = string("shares_23_cast_fp16")];
            tensor<int32, [4]> var_4972_begin_0 = const()[name = string("op_4972_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4972_end_0 = const()[name = string("op_4972_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_4972_end_mask_0 = const()[name = string("op_4972_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_4972_cast_fp16 = slice_by_index(begin = var_4972_begin_0, end = var_4972_end_0, end_mask = var_4972_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4972_cast_fp16")];
            tensor<int32, [4]> var_4977_begin_0 = const()[name = string("op_4977_begin_0"), val = tensor<int32, [4]>([11, 0, 0, 0])];
            tensor<int32, [4]> var_4977_end_0 = const()[name = string("op_4977_end_0"), val = tensor<int32, [4]>([12, 8, 528, 128])];
            tensor<bool, [4]> var_4977_end_mask_0 = const()[name = string("op_4977_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_4977_cast_fp16 = slice_by_index(begin = var_4977_begin_0, end = var_4977_end_0, end_mask = var_4977_end_mask_0, x = read_state_1)[name = string("op_4977_cast_fp16")];
            tensor<int32, [4]> var_4987_begin_0 = const()[name = string("op_4987_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_4987_end_0 = const()[name = string("op_4987_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_4987_end_mask_0 = const()[name = string("op_4987_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_4987_cast_fp16 = slice_by_index(begin = var_4987_begin_0, end = var_4987_end_0, end_mask = var_4987_end_mask_0, x = var_4977_cast_fp16)[name = string("op_4987_cast_fp16")];
            bool var_4993_transpose_x_0 = const()[name = string("op_4993_transpose_x_0"), val = bool(false)];
            bool var_4993_transpose_y_0 = const()[name = string("op_4993_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_4993_cast_fp16 = matmul(transpose_x = var_4993_transpose_x_0, transpose_y = var_4993_transpose_y_0, x = var_4972_cast_fp16, y = var_4987_cast_fp16)[name = string("op_4993_cast_fp16")];
            tensor<int32, [4]> var_4998_begin_0 = const()[name = string("op_4998_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_4998_end_0 = const()[name = string("op_4998_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_4998_end_mask_0 = const()[name = string("op_4998_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_4998_cast_fp16 = slice_by_index(begin = var_4998_begin_0, end = var_4998_end_0, end_mask = var_4998_end_mask_0, x = shares_23_cast_fp16)[name = string("op_4998_cast_fp16")];
            bool var_4999_transpose_x_0 = const()[name = string("op_4999_transpose_x_0"), val = bool(false)];
            bool var_4999_transpose_y_0 = const()[name = string("op_4999_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_23_cast_fp16 = transpose(perm = v_23_perm_0, x = var_4854_cast_fp16)[name = string("transpose_81")];
            tensor<fp16, [1, 8, 32, 128]> var_4999_cast_fp16 = matmul(transpose_x = var_4999_transpose_x_0, transpose_y = var_4999_transpose_y_0, x = var_4998_cast_fp16, y = v_23_cast_fp16)[name = string("op_4999_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_45_cast_fp16 = add(x = var_4993_cast_fp16, y = var_4999_cast_fp16)[name = string("mixed_45_cast_fp16")];
            tensor<int32, [4]> var_5016_begin_0 = const()[name = string("op_5016_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5016_end_0 = const()[name = string("op_5016_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5016_end_mask_0 = const()[name = string("op_5016_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5016_cast_fp16 = slice_by_index(begin = var_5016_begin_0, end = var_5016_end_0, end_mask = var_5016_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5016_cast_fp16")];
            tensor<int32, [4]> var_5031_begin_0 = const()[name = string("op_5031_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5031_end_0 = const()[name = string("op_5031_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5031_end_mask_0 = const()[name = string("op_5031_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5031_cast_fp16 = slice_by_index(begin = var_5031_begin_0, end = var_5031_end_0, end_mask = var_5031_end_mask_0, x = mixed_45_cast_fp16)[name = string("op_5031_cast_fp16")];
            int32 var_5033 = const()[name = string("op_5033"), val = int32(1)];
            bool mixed_47_interleave_0 = const()[name = string("mixed_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_47_cast_fp16 = concat(axis = var_5033, interleave = mixed_47_interleave_0, values = (var_5016_cast_fp16, var_5031_cast_fp16))[name = string("mixed_47_cast_fp16")];
            tensor<int32, [4]> var_5037_perm_0 = const()[name = string("op_5037_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5041 = const()[name = string("op_5041"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5037_cast_fp16 = transpose(perm = var_5037_perm_0, x = mixed_47_cast_fp16)[name = string("transpose_80")];
            tensor<fp16, [1, 16, 2048]> input_113_cast_fp16 = reshape(shape = var_5041, x = var_5037_cast_fp16)[name = string("input_113_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_11_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(488803008))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490900224))))[name = string("model_layers_11_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_80_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_self_attn_o_proj_weight_to_fp16_quantized, x = input_113_cast_fp16)[name = string("linear_80_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_143_cast_fp16 = add(x = x_133_cast_fp16, y = linear_80_cast_fp16)[name = string("x_143_cast_fp16")];
            fp16 var_5052_promoted_to_fp16 = const()[name = string("op_5052_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5054_cast_fp16 = pow(x = x_143_cast_fp16, y = var_5052_promoted_to_fp16)[name = string("op_5054_cast_fp16")];
            tensor<int32, [1]> var_5056_axes_0 = const()[name = string("op_5056_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5056_keep_dims_0 = const()[name = string("op_5056_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5056_cast_fp16 = reduce_mean(axes = var_5056_axes_0, keep_dims = var_5056_keep_dims_0, x = var_5054_cast_fp16)[name = string("op_5056_cast_fp16")];
            fp16 var_5057_to_fp16 = const()[name = string("op_5057_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5058_cast_fp16 = add(x = var_5056_cast_fp16, y = var_5057_to_fp16)[name = string("op_5058_cast_fp16")];
            fp32 var_5059_epsilon_0 = const()[name = string("op_5059_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5059_cast_fp16 = rsqrt(epsilon = var_5059_epsilon_0, x = var_5058_cast_fp16)[name = string("op_5059_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5060_cast_fp16 = mul(x = x_143_cast_fp16, y = var_5059_cast_fp16)[name = string("op_5060_cast_fp16")];
            tensor<fp16, [1024]> model_layers_11_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_11_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490902336)))];
            tensor<fp16, [1, 16, 1024]> input_115_cast_fp16 = mul(x = var_5060_cast_fp16, y = model_layers_11_post_attention_layernorm_weight_to_fp16)[name = string("input_115_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(490904448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494050240))))[name = string("model_layers_11_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_81_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_gate_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_81_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5065_cast_fp16 = silu(x = linear_81_cast_fp16)[name = string("op_5065_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_11_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(494056448))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497202240))))[name = string("model_layers_11_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_82_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_11_mlp_up_proj_weight_to_fp16_quantized, x = input_115_cast_fp16)[name = string("linear_82_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_119_cast_fp16 = mul(x = var_5065_cast_fp16, y = linear_82_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_11_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(497208448))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500354240))))[name = string("model_layers_11_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_83_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_11_mlp_down_proj_weight_to_fp16_quantized, x = input_119_cast_fp16)[name = string("linear_83_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_145_cast_fp16 = add(x = x_143_cast_fp16, y = linear_83_cast_fp16)[name = string("x_145_cast_fp16")];
            fp16 var_5079_promoted_to_fp16 = const()[name = string("op_5079_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5081_cast_fp16 = pow(x = x_145_cast_fp16, y = var_5079_promoted_to_fp16)[name = string("op_5081_cast_fp16")];
            tensor<int32, [1]> var_5083_axes_0 = const()[name = string("op_5083_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5083_keep_dims_0 = const()[name = string("op_5083_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5083_cast_fp16 = reduce_mean(axes = var_5083_axes_0, keep_dims = var_5083_keep_dims_0, x = var_5081_cast_fp16)[name = string("op_5083_cast_fp16")];
            fp16 var_5084_to_fp16 = const()[name = string("op_5084_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5085_cast_fp16 = add(x = var_5083_cast_fp16, y = var_5084_to_fp16)[name = string("op_5085_cast_fp16")];
            fp32 var_5086_epsilon_0 = const()[name = string("op_5086_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5086_cast_fp16 = rsqrt(epsilon = var_5086_epsilon_0, x = var_5085_cast_fp16)[name = string("op_5086_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5087_cast_fp16 = mul(x = x_145_cast_fp16, y = var_5086_cast_fp16)[name = string("op_5087_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500356352)))];
            tensor<fp16, [1, 16, 1024]> input_121_cast_fp16 = mul(x = var_5087_cast_fp16, y = model_layers_12_input_layernorm_weight_to_fp16)[name = string("input_121_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_12_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(500358464))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502455680))))[name = string("model_layers_12_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_84_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_12_self_attn_q_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_84_cast_fp16")];
            tensor<int32, [4]> var_5096 = const()[name = string("op_5096"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5097_cast_fp16 = reshape(shape = var_5096, x = linear_84_cast_fp16)[name = string("op_5097_cast_fp16")];
            tensor<int32, [4]> x_147_perm_0 = const()[name = string("x_147_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5105_promoted_to_fp16 = const()[name = string("op_5105_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_147_cast_fp16 = transpose(perm = x_147_perm_0, x = var_5097_cast_fp16)[name = string("transpose_79")];
            tensor<fp16, [1, 16, 16, 128]> var_5107_cast_fp16 = pow(x = x_147_cast_fp16, y = var_5105_promoted_to_fp16)[name = string("op_5107_cast_fp16")];
            tensor<int32, [1]> var_5109_axes_0 = const()[name = string("op_5109_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5109_keep_dims_0 = const()[name = string("op_5109_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5109_cast_fp16 = reduce_mean(axes = var_5109_axes_0, keep_dims = var_5109_keep_dims_0, x = var_5107_cast_fp16)[name = string("op_5109_cast_fp16")];
            fp16 var_5110_to_fp16 = const()[name = string("op_5110_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5111_cast_fp16 = add(x = var_5109_cast_fp16, y = var_5110_to_fp16)[name = string("op_5111_cast_fp16")];
            fp32 var_5112_epsilon_0 = const()[name = string("op_5112_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5112_cast_fp16 = rsqrt(epsilon = var_5112_epsilon_0, x = var_5111_cast_fp16)[name = string("op_5112_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5113_cast_fp16 = mul(x = x_147_cast_fp16, y = var_5112_cast_fp16)[name = string("op_5113_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502459840)))];
            tensor<fp16, [1, 16, 16, 128]> x_151_cast_fp16 = mul(x = var_5113_cast_fp16, y = model_layers_12_self_attn_q_norm_weight_to_fp16)[name = string("x_151_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(502460160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503508800))))[name = string("model_layers_12_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_85_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_k_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_85_cast_fp16")];
            tensor<int32, [4]> var_5122 = const()[name = string("op_5122"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5123_cast_fp16 = reshape(shape = var_5122, x = linear_85_cast_fp16)[name = string("op_5123_cast_fp16")];
            tensor<int32, [4]> x_149_perm_0 = const()[name = string("x_149_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5131_promoted_to_fp16 = const()[name = string("op_5131_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_149_cast_fp16 = transpose(perm = x_149_perm_0, x = var_5123_cast_fp16)[name = string("transpose_78")];
            tensor<fp16, [1, 8, 16, 128]> var_5133_cast_fp16 = pow(x = x_149_cast_fp16, y = var_5131_promoted_to_fp16)[name = string("op_5133_cast_fp16")];
            tensor<int32, [1]> var_5135_axes_0 = const()[name = string("op_5135_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5135_keep_dims_0 = const()[name = string("op_5135_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5135_cast_fp16 = reduce_mean(axes = var_5135_axes_0, keep_dims = var_5135_keep_dims_0, x = var_5133_cast_fp16)[name = string("op_5135_cast_fp16")];
            fp16 var_5136_to_fp16 = const()[name = string("op_5136_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5137_cast_fp16 = add(x = var_5135_cast_fp16, y = var_5136_to_fp16)[name = string("op_5137_cast_fp16")];
            fp32 var_5138_epsilon_0 = const()[name = string("op_5138_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5138_cast_fp16 = rsqrt(epsilon = var_5138_epsilon_0, x = var_5137_cast_fp16)[name = string("op_5138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5139_cast_fp16 = mul(x = x_149_cast_fp16, y = var_5138_cast_fp16)[name = string("op_5139_cast_fp16")];
            tensor<fp16, [128]> model_layers_12_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_12_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503510912)))];
            tensor<fp16, [1, 8, 16, 128]> x_153_cast_fp16 = mul(x = var_5139_cast_fp16, y = model_layers_12_self_attn_k_norm_weight_to_fp16)[name = string("x_153_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_12_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(503511232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504559872))))[name = string("model_layers_12_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_86_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_v_proj_weight_to_fp16_quantized, x = input_121_cast_fp16)[name = string("linear_86_cast_fp16")];
            tensor<int32, [4]> var_5148 = const()[name = string("op_5148"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5149_cast_fp16 = reshape(shape = var_5148, x = linear_86_cast_fp16)[name = string("op_5149_cast_fp16")];
            tensor<int32, [4]> v_25_perm_0 = const()[name = string("v_25_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_49_begin_0 = const()[name = string("first_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_49_end_0 = const()[name = string("first_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_49_end_mask_0 = const()[name = string("first_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_49_cast_fp16 = slice_by_index(begin = first_49_begin_0, end = first_49_end_0, end_mask = first_49_end_mask_0, x = x_151_cast_fp16)[name = string("first_49_cast_fp16")];
            tensor<int32, [4]> second_49_begin_0 = const()[name = string("second_49_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_49_end_0 = const()[name = string("second_49_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_49_end_mask_0 = const()[name = string("second_49_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_49_cast_fp16 = slice_by_index(begin = second_49_begin_0, end = second_49_end_0, end_mask = second_49_end_mask_0, x = x_151_cast_fp16)[name = string("second_49_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5163_cast_fp16 = mul(x = first_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5163_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5164_cast_fp16 = mul(x = second_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5164_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5166_cast_fp16 = sub(x = var_5163_cast_fp16, y = var_5164_cast_fp16)[name = string("op_5166_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5167_cast_fp16 = mul(x = second_49_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5167_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5168_cast_fp16 = mul(x = first_49_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5168_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5170_cast_fp16 = add(x = var_5167_cast_fp16, y = var_5168_cast_fp16)[name = string("op_5170_cast_fp16")];
            int32 var_5172 = const()[name = string("op_5172"), val = int32(-1)];
            bool q_25_interleave_0 = const()[name = string("q_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_25_cast_fp16 = concat(axis = var_5172, interleave = q_25_interleave_0, values = (var_5166_cast_fp16, var_5170_cast_fp16))[name = string("q_25_cast_fp16")];
            tensor<int32, [4]> first_51_begin_0 = const()[name = string("first_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_51_end_0 = const()[name = string("first_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_51_end_mask_0 = const()[name = string("first_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_51_cast_fp16 = slice_by_index(begin = first_51_begin_0, end = first_51_end_0, end_mask = first_51_end_mask_0, x = x_153_cast_fp16)[name = string("first_51_cast_fp16")];
            tensor<int32, [4]> second_51_begin_0 = const()[name = string("second_51_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_51_end_0 = const()[name = string("second_51_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_51_end_mask_0 = const()[name = string("second_51_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_51_cast_fp16 = slice_by_index(begin = second_51_begin_0, end = second_51_end_0, end_mask = second_51_end_mask_0, x = x_153_cast_fp16)[name = string("second_51_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5184_cast_fp16 = mul(x = first_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5184_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5185_cast_fp16 = mul(x = second_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5185_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5187_cast_fp16 = sub(x = var_5184_cast_fp16, y = var_5185_cast_fp16)[name = string("op_5187_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5188_cast_fp16 = mul(x = second_51_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5188_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5189_cast_fp16 = mul(x = first_51_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5189_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5191_cast_fp16 = add(x = var_5188_cast_fp16, y = var_5189_cast_fp16)[name = string("op_5191_cast_fp16")];
            int32 var_5193 = const()[name = string("op_5193"), val = int32(-1)];
            bool k_25_interleave_0 = const()[name = string("k_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_25_cast_fp16 = concat(axis = var_5193, interleave = k_25_interleave_0, values = (var_5187_cast_fp16, var_5191_cast_fp16))[name = string("k_25_cast_fp16")];
            tensor<int32, [4]> var_5204_begin_0 = const()[name = string("op_5204_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5204_end_0 = const()[name = string("op_5204_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5204_end_mask_0 = const()[name = string("op_5204_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5204_stride_0 = const()[name = string("op_5204_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5204_cast_fp16 = slice_by_index(begin = var_5204_begin_0, end = var_5204_end_0, end_mask = var_5204_end_mask_0, stride = var_5204_stride_0, x = q_25_cast_fp16)[name = string("op_5204_cast_fp16")];
            tensor<int32, [4]> var_5214_begin_0 = const()[name = string("op_5214_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5214_end_0 = const()[name = string("op_5214_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5214_end_mask_0 = const()[name = string("op_5214_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5214_stride_0 = const()[name = string("op_5214_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5214_cast_fp16 = slice_by_index(begin = var_5214_begin_0, end = var_5214_end_0, end_mask = var_5214_end_mask_0, stride = var_5214_stride_0, x = q_25_cast_fp16)[name = string("op_5214_cast_fp16")];
            int32 var_5216 = const()[name = string("op_5216"), val = int32(2)];
            bool stacked_25_interleave_0 = const()[name = string("stacked_25_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_25_cast_fp16 = concat(axis = var_5216, interleave = stacked_25_interleave_0, values = (var_5204_cast_fp16, var_5214_cast_fp16))[name = string("stacked_25_cast_fp16")];
            tensor<int32, [4]> var_5222_begin_0 = const()[name = string("op_5222_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5222_end_0 = const()[name = string("op_5222_end_0"), val = tensor<int32, [4]>([13, 8, 528, 128])];
            tensor<bool, [4]> var_5222_end_mask_0 = const()[name = string("op_5222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5222_cast_fp16 = slice_by_index(begin = var_5222_begin_0, end = var_5222_end_0, end_mask = var_5222_end_mask_0, x = read_state_0)[name = string("op_5222_cast_fp16")];
            tensor<int32, [4]> var_5232_begin_0 = const()[name = string("op_5232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5232_end_0 = const()[name = string("op_5232_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5232_end_mask_0 = const()[name = string("op_5232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5232_cast_fp16 = slice_by_index(begin = var_5232_begin_0, end = var_5232_end_0, end_mask = var_5232_end_mask_0, x = var_5222_cast_fp16)[name = string("op_5232_cast_fp16")];
            bool var_5241_transpose_x_1 = const()[name = string("op_5241_transpose_x_1"), val = bool(false)];
            bool var_5241_transpose_y_1 = const()[name = string("op_5241_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5241_cast_fp16 = matmul(transpose_x = var_5241_transpose_x_1, transpose_y = var_5241_transpose_y_1, x = var_5232_cast_fp16, y = stacked_25_cast_fp16)[name = string("op_5241_cast_fp16")];
            tensor<int32, [4]> var_5244_perm_0 = const()[name = string("op_5244_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5245_to_fp16 = const()[name = string("op_5245_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5244_cast_fp16 = transpose(perm = var_5244_perm_0, x = var_5241_cast_fp16)[name = string("transpose_77")];
            tensor<fp16, [1, 8, 32, 512]> var_5246_cast_fp16 = mul(x = var_5244_cast_fp16, y = var_5245_to_fp16)[name = string("op_5246_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_25_cast_fp16 = add(x = var_5246_cast_fp16, y = cacheMask)[name = string("cache_weights_25_cast_fp16")];
            bool var_5252_transpose_x_1 = const()[name = string("op_5252_transpose_x_1"), val = bool(false)];
            bool var_5252_transpose_y_1 = const()[name = string("op_5252_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5252_cast_fp16 = matmul(transpose_x = var_5252_transpose_x_1, transpose_y = var_5252_transpose_y_1, x = stacked_25_cast_fp16, y = k_25_cast_fp16)[name = string("op_5252_cast_fp16")];
            fp16 var_5253_to_fp16 = const()[name = string("op_5253_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5254_cast_fp16 = mul(x = var_5252_cast_fp16, y = var_5253_to_fp16)[name = string("op_5254_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_25_cast_fp16 = add(x = var_5254_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_25_cast_fp16")];
            int32 var_5258 = const()[name = string("op_5258"), val = int32(-1)];
            bool var_5259_interleave_0 = const()[name = string("op_5259_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5259_cast_fp16 = concat(axis = var_5258, interleave = var_5259_interleave_0, values = (cache_weights_25_cast_fp16, call_weights_25_cast_fp16))[name = string("op_5259_cast_fp16")];
            int32 var_5260 = const()[name = string("op_5260"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_25_cast_fp16 = softmax(axis = var_5260, x = var_5259_cast_fp16)[name = string("shares_25_cast_fp16")];
            tensor<int32, [4]> var_5267_begin_0 = const()[name = string("op_5267_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5267_end_0 = const()[name = string("op_5267_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5267_end_mask_0 = const()[name = string("op_5267_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5267_cast_fp16 = slice_by_index(begin = var_5267_begin_0, end = var_5267_end_0, end_mask = var_5267_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5267_cast_fp16")];
            tensor<int32, [4]> var_5272_begin_0 = const()[name = string("op_5272_begin_0"), val = tensor<int32, [4]>([12, 0, 0, 0])];
            tensor<int32, [4]> var_5272_end_0 = const()[name = string("op_5272_end_0"), val = tensor<int32, [4]>([13, 8, 528, 128])];
            tensor<bool, [4]> var_5272_end_mask_0 = const()[name = string("op_5272_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5272_cast_fp16 = slice_by_index(begin = var_5272_begin_0, end = var_5272_end_0, end_mask = var_5272_end_mask_0, x = read_state_1)[name = string("op_5272_cast_fp16")];
            tensor<int32, [4]> var_5282_begin_0 = const()[name = string("op_5282_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5282_end_0 = const()[name = string("op_5282_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5282_end_mask_0 = const()[name = string("op_5282_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5282_cast_fp16 = slice_by_index(begin = var_5282_begin_0, end = var_5282_end_0, end_mask = var_5282_end_mask_0, x = var_5272_cast_fp16)[name = string("op_5282_cast_fp16")];
            bool var_5288_transpose_x_0 = const()[name = string("op_5288_transpose_x_0"), val = bool(false)];
            bool var_5288_transpose_y_0 = const()[name = string("op_5288_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5288_cast_fp16 = matmul(transpose_x = var_5288_transpose_x_0, transpose_y = var_5288_transpose_y_0, x = var_5267_cast_fp16, y = var_5282_cast_fp16)[name = string("op_5288_cast_fp16")];
            tensor<int32, [4]> var_5293_begin_0 = const()[name = string("op_5293_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5293_end_0 = const()[name = string("op_5293_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5293_end_mask_0 = const()[name = string("op_5293_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5293_cast_fp16 = slice_by_index(begin = var_5293_begin_0, end = var_5293_end_0, end_mask = var_5293_end_mask_0, x = shares_25_cast_fp16)[name = string("op_5293_cast_fp16")];
            bool var_5294_transpose_x_0 = const()[name = string("op_5294_transpose_x_0"), val = bool(false)];
            bool var_5294_transpose_y_0 = const()[name = string("op_5294_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_25_cast_fp16 = transpose(perm = v_25_perm_0, x = var_5149_cast_fp16)[name = string("transpose_76")];
            tensor<fp16, [1, 8, 32, 128]> var_5294_cast_fp16 = matmul(transpose_x = var_5294_transpose_x_0, transpose_y = var_5294_transpose_y_0, x = var_5293_cast_fp16, y = v_25_cast_fp16)[name = string("op_5294_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_49_cast_fp16 = add(x = var_5288_cast_fp16, y = var_5294_cast_fp16)[name = string("mixed_49_cast_fp16")];
            tensor<int32, [4]> var_5311_begin_0 = const()[name = string("op_5311_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5311_end_0 = const()[name = string("op_5311_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5311_end_mask_0 = const()[name = string("op_5311_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5311_cast_fp16 = slice_by_index(begin = var_5311_begin_0, end = var_5311_end_0, end_mask = var_5311_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5311_cast_fp16")];
            tensor<int32, [4]> var_5326_begin_0 = const()[name = string("op_5326_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5326_end_0 = const()[name = string("op_5326_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5326_end_mask_0 = const()[name = string("op_5326_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5326_cast_fp16 = slice_by_index(begin = var_5326_begin_0, end = var_5326_end_0, end_mask = var_5326_end_mask_0, x = mixed_49_cast_fp16)[name = string("op_5326_cast_fp16")];
            int32 var_5328 = const()[name = string("op_5328"), val = int32(1)];
            bool mixed_51_interleave_0 = const()[name = string("mixed_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_51_cast_fp16 = concat(axis = var_5328, interleave = mixed_51_interleave_0, values = (var_5311_cast_fp16, var_5326_cast_fp16))[name = string("mixed_51_cast_fp16")];
            tensor<int32, [4]> var_5332_perm_0 = const()[name = string("op_5332_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5336 = const()[name = string("op_5336"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5332_cast_fp16 = transpose(perm = var_5332_perm_0, x = mixed_51_cast_fp16)[name = string("transpose_75")];
            tensor<fp16, [1, 16, 2048]> input_123_cast_fp16 = reshape(shape = var_5336, x = var_5332_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_12_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(504561984))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506659200))))[name = string("model_layers_12_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_87_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_self_attn_o_proj_weight_to_fp16_quantized, x = input_123_cast_fp16)[name = string("linear_87_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_155_cast_fp16 = add(x = x_145_cast_fp16, y = linear_87_cast_fp16)[name = string("x_155_cast_fp16")];
            fp16 var_5347_promoted_to_fp16 = const()[name = string("op_5347_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5349_cast_fp16 = pow(x = x_155_cast_fp16, y = var_5347_promoted_to_fp16)[name = string("op_5349_cast_fp16")];
            tensor<int32, [1]> var_5351_axes_0 = const()[name = string("op_5351_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5351_keep_dims_0 = const()[name = string("op_5351_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5351_cast_fp16 = reduce_mean(axes = var_5351_axes_0, keep_dims = var_5351_keep_dims_0, x = var_5349_cast_fp16)[name = string("op_5351_cast_fp16")];
            fp16 var_5352_to_fp16 = const()[name = string("op_5352_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5353_cast_fp16 = add(x = var_5351_cast_fp16, y = var_5352_to_fp16)[name = string("op_5353_cast_fp16")];
            fp32 var_5354_epsilon_0 = const()[name = string("op_5354_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5354_cast_fp16 = rsqrt(epsilon = var_5354_epsilon_0, x = var_5353_cast_fp16)[name = string("op_5354_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5355_cast_fp16 = mul(x = x_155_cast_fp16, y = var_5354_cast_fp16)[name = string("op_5355_cast_fp16")];
            tensor<fp16, [1024]> model_layers_12_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_12_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506661312)))];
            tensor<fp16, [1, 16, 1024]> input_125_cast_fp16 = mul(x = var_5355_cast_fp16, y = model_layers_12_post_attention_layernorm_weight_to_fp16)[name = string("input_125_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(506663424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509809216))))[name = string("model_layers_12_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_88_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_gate_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_88_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5360_cast_fp16 = silu(x = linear_88_cast_fp16)[name = string("op_5360_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_12_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(509815424))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512961216))))[name = string("model_layers_12_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_89_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_12_mlp_up_proj_weight_to_fp16_quantized, x = input_125_cast_fp16)[name = string("linear_89_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_129_cast_fp16 = mul(x = var_5360_cast_fp16, y = linear_89_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_12_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(512967424))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516113216))))[name = string("model_layers_12_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_90_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_12_mlp_down_proj_weight_to_fp16_quantized, x = input_129_cast_fp16)[name = string("linear_90_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_157_cast_fp16 = add(x = x_155_cast_fp16, y = linear_90_cast_fp16)[name = string("x_157_cast_fp16")];
            fp16 var_5374_promoted_to_fp16 = const()[name = string("op_5374_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5376_cast_fp16 = pow(x = x_157_cast_fp16, y = var_5374_promoted_to_fp16)[name = string("op_5376_cast_fp16")];
            tensor<int32, [1]> var_5378_axes_0 = const()[name = string("op_5378_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5378_keep_dims_0 = const()[name = string("op_5378_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5378_cast_fp16 = reduce_mean(axes = var_5378_axes_0, keep_dims = var_5378_keep_dims_0, x = var_5376_cast_fp16)[name = string("op_5378_cast_fp16")];
            fp16 var_5379_to_fp16 = const()[name = string("op_5379_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5380_cast_fp16 = add(x = var_5378_cast_fp16, y = var_5379_to_fp16)[name = string("op_5380_cast_fp16")];
            fp32 var_5381_epsilon_0 = const()[name = string("op_5381_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5381_cast_fp16 = rsqrt(epsilon = var_5381_epsilon_0, x = var_5380_cast_fp16)[name = string("op_5381_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5382_cast_fp16 = mul(x = x_157_cast_fp16, y = var_5381_cast_fp16)[name = string("op_5382_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516115328)))];
            tensor<fp16, [1, 16, 1024]> input_131_cast_fp16 = mul(x = var_5382_cast_fp16, y = model_layers_13_input_layernorm_weight_to_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_13_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(516117440))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518214656))))[name = string("model_layers_13_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_91_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_13_self_attn_q_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_91_cast_fp16")];
            tensor<int32, [4]> var_5391 = const()[name = string("op_5391"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5392_cast_fp16 = reshape(shape = var_5391, x = linear_91_cast_fp16)[name = string("op_5392_cast_fp16")];
            tensor<int32, [4]> x_159_perm_0 = const()[name = string("x_159_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5400_promoted_to_fp16 = const()[name = string("op_5400_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_159_cast_fp16 = transpose(perm = x_159_perm_0, x = var_5392_cast_fp16)[name = string("transpose_74")];
            tensor<fp16, [1, 16, 16, 128]> var_5402_cast_fp16 = pow(x = x_159_cast_fp16, y = var_5400_promoted_to_fp16)[name = string("op_5402_cast_fp16")];
            tensor<int32, [1]> var_5404_axes_0 = const()[name = string("op_5404_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5404_keep_dims_0 = const()[name = string("op_5404_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5404_cast_fp16 = reduce_mean(axes = var_5404_axes_0, keep_dims = var_5404_keep_dims_0, x = var_5402_cast_fp16)[name = string("op_5404_cast_fp16")];
            fp16 var_5405_to_fp16 = const()[name = string("op_5405_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5406_cast_fp16 = add(x = var_5404_cast_fp16, y = var_5405_to_fp16)[name = string("op_5406_cast_fp16")];
            fp32 var_5407_epsilon_0 = const()[name = string("op_5407_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5407_cast_fp16 = rsqrt(epsilon = var_5407_epsilon_0, x = var_5406_cast_fp16)[name = string("op_5407_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5408_cast_fp16 = mul(x = x_159_cast_fp16, y = var_5407_cast_fp16)[name = string("op_5408_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518218816)))];
            tensor<fp16, [1, 16, 16, 128]> x_163_cast_fp16 = mul(x = var_5408_cast_fp16, y = model_layers_13_self_attn_q_norm_weight_to_fp16)[name = string("x_163_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(518219136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519267776))))[name = string("model_layers_13_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_92_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_k_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_92_cast_fp16")];
            tensor<int32, [4]> var_5417 = const()[name = string("op_5417"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5418_cast_fp16 = reshape(shape = var_5417, x = linear_92_cast_fp16)[name = string("op_5418_cast_fp16")];
            tensor<int32, [4]> x_161_perm_0 = const()[name = string("x_161_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5426_promoted_to_fp16 = const()[name = string("op_5426_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_161_cast_fp16 = transpose(perm = x_161_perm_0, x = var_5418_cast_fp16)[name = string("transpose_73")];
            tensor<fp16, [1, 8, 16, 128]> var_5428_cast_fp16 = pow(x = x_161_cast_fp16, y = var_5426_promoted_to_fp16)[name = string("op_5428_cast_fp16")];
            tensor<int32, [1]> var_5430_axes_0 = const()[name = string("op_5430_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5430_keep_dims_0 = const()[name = string("op_5430_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5430_cast_fp16 = reduce_mean(axes = var_5430_axes_0, keep_dims = var_5430_keep_dims_0, x = var_5428_cast_fp16)[name = string("op_5430_cast_fp16")];
            fp16 var_5431_to_fp16 = const()[name = string("op_5431_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5432_cast_fp16 = add(x = var_5430_cast_fp16, y = var_5431_to_fp16)[name = string("op_5432_cast_fp16")];
            fp32 var_5433_epsilon_0 = const()[name = string("op_5433_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5433_cast_fp16 = rsqrt(epsilon = var_5433_epsilon_0, x = var_5432_cast_fp16)[name = string("op_5433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5434_cast_fp16 = mul(x = x_161_cast_fp16, y = var_5433_cast_fp16)[name = string("op_5434_cast_fp16")];
            tensor<fp16, [128]> model_layers_13_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_13_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519269888)))];
            tensor<fp16, [1, 8, 16, 128]> x_165_cast_fp16 = mul(x = var_5434_cast_fp16, y = model_layers_13_self_attn_k_norm_weight_to_fp16)[name = string("x_165_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_13_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(519270208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520318848))))[name = string("model_layers_13_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_93_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_v_proj_weight_to_fp16_quantized, x = input_131_cast_fp16)[name = string("linear_93_cast_fp16")];
            tensor<int32, [4]> var_5443 = const()[name = string("op_5443"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5444_cast_fp16 = reshape(shape = var_5443, x = linear_93_cast_fp16)[name = string("op_5444_cast_fp16")];
            tensor<int32, [4]> v_27_perm_0 = const()[name = string("v_27_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_53_begin_0 = const()[name = string("first_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_53_end_0 = const()[name = string("first_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_53_end_mask_0 = const()[name = string("first_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_53_cast_fp16 = slice_by_index(begin = first_53_begin_0, end = first_53_end_0, end_mask = first_53_end_mask_0, x = x_163_cast_fp16)[name = string("first_53_cast_fp16")];
            tensor<int32, [4]> second_53_begin_0 = const()[name = string("second_53_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_53_end_0 = const()[name = string("second_53_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_53_end_mask_0 = const()[name = string("second_53_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_53_cast_fp16 = slice_by_index(begin = second_53_begin_0, end = second_53_end_0, end_mask = second_53_end_mask_0, x = x_163_cast_fp16)[name = string("second_53_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5458_cast_fp16 = mul(x = first_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5458_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5459_cast_fp16 = mul(x = second_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5459_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5461_cast_fp16 = sub(x = var_5458_cast_fp16, y = var_5459_cast_fp16)[name = string("op_5461_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5462_cast_fp16 = mul(x = second_53_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5462_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5463_cast_fp16 = mul(x = first_53_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5463_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5465_cast_fp16 = add(x = var_5462_cast_fp16, y = var_5463_cast_fp16)[name = string("op_5465_cast_fp16")];
            int32 var_5467 = const()[name = string("op_5467"), val = int32(-1)];
            bool q_27_interleave_0 = const()[name = string("q_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_27_cast_fp16 = concat(axis = var_5467, interleave = q_27_interleave_0, values = (var_5461_cast_fp16, var_5465_cast_fp16))[name = string("q_27_cast_fp16")];
            tensor<int32, [4]> first_55_begin_0 = const()[name = string("first_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_55_end_0 = const()[name = string("first_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_55_end_mask_0 = const()[name = string("first_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_55_cast_fp16 = slice_by_index(begin = first_55_begin_0, end = first_55_end_0, end_mask = first_55_end_mask_0, x = x_165_cast_fp16)[name = string("first_55_cast_fp16")];
            tensor<int32, [4]> second_55_begin_0 = const()[name = string("second_55_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_55_end_0 = const()[name = string("second_55_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_55_end_mask_0 = const()[name = string("second_55_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_55_cast_fp16 = slice_by_index(begin = second_55_begin_0, end = second_55_end_0, end_mask = second_55_end_mask_0, x = x_165_cast_fp16)[name = string("second_55_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5479_cast_fp16 = mul(x = first_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5479_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5480_cast_fp16 = mul(x = second_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5480_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5482_cast_fp16 = sub(x = var_5479_cast_fp16, y = var_5480_cast_fp16)[name = string("op_5482_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5483_cast_fp16 = mul(x = second_55_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5483_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5484_cast_fp16 = mul(x = first_55_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5484_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5486_cast_fp16 = add(x = var_5483_cast_fp16, y = var_5484_cast_fp16)[name = string("op_5486_cast_fp16")];
            int32 var_5488 = const()[name = string("op_5488"), val = int32(-1)];
            bool k_27_interleave_0 = const()[name = string("k_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_27_cast_fp16 = concat(axis = var_5488, interleave = k_27_interleave_0, values = (var_5482_cast_fp16, var_5486_cast_fp16))[name = string("k_27_cast_fp16")];
            tensor<int32, [4]> var_5499_begin_0 = const()[name = string("op_5499_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5499_end_0 = const()[name = string("op_5499_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5499_end_mask_0 = const()[name = string("op_5499_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5499_stride_0 = const()[name = string("op_5499_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5499_cast_fp16 = slice_by_index(begin = var_5499_begin_0, end = var_5499_end_0, end_mask = var_5499_end_mask_0, stride = var_5499_stride_0, x = q_27_cast_fp16)[name = string("op_5499_cast_fp16")];
            tensor<int32, [4]> var_5509_begin_0 = const()[name = string("op_5509_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5509_end_0 = const()[name = string("op_5509_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5509_end_mask_0 = const()[name = string("op_5509_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5509_stride_0 = const()[name = string("op_5509_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5509_cast_fp16 = slice_by_index(begin = var_5509_begin_0, end = var_5509_end_0, end_mask = var_5509_end_mask_0, stride = var_5509_stride_0, x = q_27_cast_fp16)[name = string("op_5509_cast_fp16")];
            int32 var_5511 = const()[name = string("op_5511"), val = int32(2)];
            bool stacked_27_interleave_0 = const()[name = string("stacked_27_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_27_cast_fp16 = concat(axis = var_5511, interleave = stacked_27_interleave_0, values = (var_5499_cast_fp16, var_5509_cast_fp16))[name = string("stacked_27_cast_fp16")];
            tensor<int32, [4]> var_5517_begin_0 = const()[name = string("op_5517_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5517_end_0 = const()[name = string("op_5517_end_0"), val = tensor<int32, [4]>([14, 8, 528, 128])];
            tensor<bool, [4]> var_5517_end_mask_0 = const()[name = string("op_5517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5517_cast_fp16 = slice_by_index(begin = var_5517_begin_0, end = var_5517_end_0, end_mask = var_5517_end_mask_0, x = read_state_0)[name = string("op_5517_cast_fp16")];
            tensor<int32, [4]> var_5527_begin_0 = const()[name = string("op_5527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5527_end_0 = const()[name = string("op_5527_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5527_end_mask_0 = const()[name = string("op_5527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5527_cast_fp16 = slice_by_index(begin = var_5527_begin_0, end = var_5527_end_0, end_mask = var_5527_end_mask_0, x = var_5517_cast_fp16)[name = string("op_5527_cast_fp16")];
            bool var_5536_transpose_x_1 = const()[name = string("op_5536_transpose_x_1"), val = bool(false)];
            bool var_5536_transpose_y_1 = const()[name = string("op_5536_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5536_cast_fp16 = matmul(transpose_x = var_5536_transpose_x_1, transpose_y = var_5536_transpose_y_1, x = var_5527_cast_fp16, y = stacked_27_cast_fp16)[name = string("op_5536_cast_fp16")];
            tensor<int32, [4]> var_5539_perm_0 = const()[name = string("op_5539_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5540_to_fp16 = const()[name = string("op_5540_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5539_cast_fp16 = transpose(perm = var_5539_perm_0, x = var_5536_cast_fp16)[name = string("transpose_72")];
            tensor<fp16, [1, 8, 32, 512]> var_5541_cast_fp16 = mul(x = var_5539_cast_fp16, y = var_5540_to_fp16)[name = string("op_5541_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_27_cast_fp16 = add(x = var_5541_cast_fp16, y = cacheMask)[name = string("cache_weights_27_cast_fp16")];
            bool var_5547_transpose_x_1 = const()[name = string("op_5547_transpose_x_1"), val = bool(false)];
            bool var_5547_transpose_y_1 = const()[name = string("op_5547_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5547_cast_fp16 = matmul(transpose_x = var_5547_transpose_x_1, transpose_y = var_5547_transpose_y_1, x = stacked_27_cast_fp16, y = k_27_cast_fp16)[name = string("op_5547_cast_fp16")];
            fp16 var_5548_to_fp16 = const()[name = string("op_5548_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5549_cast_fp16 = mul(x = var_5547_cast_fp16, y = var_5548_to_fp16)[name = string("op_5549_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_27_cast_fp16 = add(x = var_5549_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_27_cast_fp16")];
            int32 var_5553 = const()[name = string("op_5553"), val = int32(-1)];
            bool var_5554_interleave_0 = const()[name = string("op_5554_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5554_cast_fp16 = concat(axis = var_5553, interleave = var_5554_interleave_0, values = (cache_weights_27_cast_fp16, call_weights_27_cast_fp16))[name = string("op_5554_cast_fp16")];
            int32 var_5555 = const()[name = string("op_5555"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_27_cast_fp16 = softmax(axis = var_5555, x = var_5554_cast_fp16)[name = string("shares_27_cast_fp16")];
            tensor<int32, [4]> var_5562_begin_0 = const()[name = string("op_5562_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5562_end_0 = const()[name = string("op_5562_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5562_end_mask_0 = const()[name = string("op_5562_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5562_cast_fp16 = slice_by_index(begin = var_5562_begin_0, end = var_5562_end_0, end_mask = var_5562_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5562_cast_fp16")];
            tensor<int32, [4]> var_5567_begin_0 = const()[name = string("op_5567_begin_0"), val = tensor<int32, [4]>([13, 0, 0, 0])];
            tensor<int32, [4]> var_5567_end_0 = const()[name = string("op_5567_end_0"), val = tensor<int32, [4]>([14, 8, 528, 128])];
            tensor<bool, [4]> var_5567_end_mask_0 = const()[name = string("op_5567_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5567_cast_fp16 = slice_by_index(begin = var_5567_begin_0, end = var_5567_end_0, end_mask = var_5567_end_mask_0, x = read_state_1)[name = string("op_5567_cast_fp16")];
            tensor<int32, [4]> var_5577_begin_0 = const()[name = string("op_5577_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5577_end_0 = const()[name = string("op_5577_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5577_end_mask_0 = const()[name = string("op_5577_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5577_cast_fp16 = slice_by_index(begin = var_5577_begin_0, end = var_5577_end_0, end_mask = var_5577_end_mask_0, x = var_5567_cast_fp16)[name = string("op_5577_cast_fp16")];
            bool var_5583_transpose_x_0 = const()[name = string("op_5583_transpose_x_0"), val = bool(false)];
            bool var_5583_transpose_y_0 = const()[name = string("op_5583_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5583_cast_fp16 = matmul(transpose_x = var_5583_transpose_x_0, transpose_y = var_5583_transpose_y_0, x = var_5562_cast_fp16, y = var_5577_cast_fp16)[name = string("op_5583_cast_fp16")];
            tensor<int32, [4]> var_5588_begin_0 = const()[name = string("op_5588_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5588_end_0 = const()[name = string("op_5588_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5588_end_mask_0 = const()[name = string("op_5588_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5588_cast_fp16 = slice_by_index(begin = var_5588_begin_0, end = var_5588_end_0, end_mask = var_5588_end_mask_0, x = shares_27_cast_fp16)[name = string("op_5588_cast_fp16")];
            bool var_5589_transpose_x_0 = const()[name = string("op_5589_transpose_x_0"), val = bool(false)];
            bool var_5589_transpose_y_0 = const()[name = string("op_5589_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_27_cast_fp16 = transpose(perm = v_27_perm_0, x = var_5444_cast_fp16)[name = string("transpose_71")];
            tensor<fp16, [1, 8, 32, 128]> var_5589_cast_fp16 = matmul(transpose_x = var_5589_transpose_x_0, transpose_y = var_5589_transpose_y_0, x = var_5588_cast_fp16, y = v_27_cast_fp16)[name = string("op_5589_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_53_cast_fp16 = add(x = var_5583_cast_fp16, y = var_5589_cast_fp16)[name = string("mixed_53_cast_fp16")];
            tensor<int32, [4]> var_5606_begin_0 = const()[name = string("op_5606_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5606_end_0 = const()[name = string("op_5606_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5606_end_mask_0 = const()[name = string("op_5606_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5606_cast_fp16 = slice_by_index(begin = var_5606_begin_0, end = var_5606_end_0, end_mask = var_5606_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5606_cast_fp16")];
            tensor<int32, [4]> var_5621_begin_0 = const()[name = string("op_5621_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5621_end_0 = const()[name = string("op_5621_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5621_end_mask_0 = const()[name = string("op_5621_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5621_cast_fp16 = slice_by_index(begin = var_5621_begin_0, end = var_5621_end_0, end_mask = var_5621_end_mask_0, x = mixed_53_cast_fp16)[name = string("op_5621_cast_fp16")];
            int32 var_5623 = const()[name = string("op_5623"), val = int32(1)];
            bool mixed_55_interleave_0 = const()[name = string("mixed_55_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_55_cast_fp16 = concat(axis = var_5623, interleave = mixed_55_interleave_0, values = (var_5606_cast_fp16, var_5621_cast_fp16))[name = string("mixed_55_cast_fp16")];
            tensor<int32, [4]> var_5627_perm_0 = const()[name = string("op_5627_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5631 = const()[name = string("op_5631"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5627_cast_fp16 = transpose(perm = var_5627_perm_0, x = mixed_55_cast_fp16)[name = string("transpose_70")];
            tensor<fp16, [1, 16, 2048]> input_133_cast_fp16 = reshape(shape = var_5631, x = var_5627_cast_fp16)[name = string("input_133_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_13_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(520320960))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522418176))))[name = string("model_layers_13_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_94_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_self_attn_o_proj_weight_to_fp16_quantized, x = input_133_cast_fp16)[name = string("linear_94_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_167_cast_fp16 = add(x = x_157_cast_fp16, y = linear_94_cast_fp16)[name = string("x_167_cast_fp16")];
            fp16 var_5642_promoted_to_fp16 = const()[name = string("op_5642_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5644_cast_fp16 = pow(x = x_167_cast_fp16, y = var_5642_promoted_to_fp16)[name = string("op_5644_cast_fp16")];
            tensor<int32, [1]> var_5646_axes_0 = const()[name = string("op_5646_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5646_keep_dims_0 = const()[name = string("op_5646_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5646_cast_fp16 = reduce_mean(axes = var_5646_axes_0, keep_dims = var_5646_keep_dims_0, x = var_5644_cast_fp16)[name = string("op_5646_cast_fp16")];
            fp16 var_5647_to_fp16 = const()[name = string("op_5647_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5648_cast_fp16 = add(x = var_5646_cast_fp16, y = var_5647_to_fp16)[name = string("op_5648_cast_fp16")];
            fp32 var_5649_epsilon_0 = const()[name = string("op_5649_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5649_cast_fp16 = rsqrt(epsilon = var_5649_epsilon_0, x = var_5648_cast_fp16)[name = string("op_5649_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5650_cast_fp16 = mul(x = x_167_cast_fp16, y = var_5649_cast_fp16)[name = string("op_5650_cast_fp16")];
            tensor<fp16, [1024]> model_layers_13_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_13_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522420288)))];
            tensor<fp16, [1, 16, 1024]> input_135_cast_fp16 = mul(x = var_5650_cast_fp16, y = model_layers_13_post_attention_layernorm_weight_to_fp16)[name = string("input_135_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(522422400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525568192))))[name = string("model_layers_13_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_95_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_gate_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_95_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5655_cast_fp16 = silu(x = linear_95_cast_fp16)[name = string("op_5655_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_13_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(525574400))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528720192))))[name = string("model_layers_13_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_96_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_13_mlp_up_proj_weight_to_fp16_quantized, x = input_135_cast_fp16)[name = string("linear_96_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_139_cast_fp16 = mul(x = var_5655_cast_fp16, y = linear_96_cast_fp16)[name = string("input_139_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_13_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(528726400))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531872192))))[name = string("model_layers_13_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_97_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_13_mlp_down_proj_weight_to_fp16_quantized, x = input_139_cast_fp16)[name = string("linear_97_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_169_cast_fp16 = add(x = x_167_cast_fp16, y = linear_97_cast_fp16)[name = string("x_169_cast_fp16")];
            fp16 var_5669_promoted_to_fp16 = const()[name = string("op_5669_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5671_cast_fp16 = pow(x = x_169_cast_fp16, y = var_5669_promoted_to_fp16)[name = string("op_5671_cast_fp16")];
            tensor<int32, [1]> var_5673_axes_0 = const()[name = string("op_5673_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5673_keep_dims_0 = const()[name = string("op_5673_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5673_cast_fp16 = reduce_mean(axes = var_5673_axes_0, keep_dims = var_5673_keep_dims_0, x = var_5671_cast_fp16)[name = string("op_5673_cast_fp16")];
            fp16 var_5674_to_fp16 = const()[name = string("op_5674_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5675_cast_fp16 = add(x = var_5673_cast_fp16, y = var_5674_to_fp16)[name = string("op_5675_cast_fp16")];
            fp32 var_5676_epsilon_0 = const()[name = string("op_5676_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5676_cast_fp16 = rsqrt(epsilon = var_5676_epsilon_0, x = var_5675_cast_fp16)[name = string("op_5676_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5677_cast_fp16 = mul(x = x_169_cast_fp16, y = var_5676_cast_fp16)[name = string("op_5677_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531874304)))];
            tensor<fp16, [1, 16, 1024]> input_141_cast_fp16 = mul(x = var_5677_cast_fp16, y = model_layers_14_input_layernorm_weight_to_fp16)[name = string("input_141_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_14_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(531876416))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533973632))))[name = string("model_layers_14_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_98_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_14_self_attn_q_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_98_cast_fp16")];
            tensor<int32, [4]> var_5686 = const()[name = string("op_5686"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5687_cast_fp16 = reshape(shape = var_5686, x = linear_98_cast_fp16)[name = string("op_5687_cast_fp16")];
            tensor<int32, [4]> x_171_perm_0 = const()[name = string("x_171_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5695_promoted_to_fp16 = const()[name = string("op_5695_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_171_cast_fp16 = transpose(perm = x_171_perm_0, x = var_5687_cast_fp16)[name = string("transpose_69")];
            tensor<fp16, [1, 16, 16, 128]> var_5697_cast_fp16 = pow(x = x_171_cast_fp16, y = var_5695_promoted_to_fp16)[name = string("op_5697_cast_fp16")];
            tensor<int32, [1]> var_5699_axes_0 = const()[name = string("op_5699_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5699_keep_dims_0 = const()[name = string("op_5699_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5699_cast_fp16 = reduce_mean(axes = var_5699_axes_0, keep_dims = var_5699_keep_dims_0, x = var_5697_cast_fp16)[name = string("op_5699_cast_fp16")];
            fp16 var_5700_to_fp16 = const()[name = string("op_5700_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5701_cast_fp16 = add(x = var_5699_cast_fp16, y = var_5700_to_fp16)[name = string("op_5701_cast_fp16")];
            fp32 var_5702_epsilon_0 = const()[name = string("op_5702_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5702_cast_fp16 = rsqrt(epsilon = var_5702_epsilon_0, x = var_5701_cast_fp16)[name = string("op_5702_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5703_cast_fp16 = mul(x = x_171_cast_fp16, y = var_5702_cast_fp16)[name = string("op_5703_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533977792)))];
            tensor<fp16, [1, 16, 16, 128]> x_175_cast_fp16 = mul(x = var_5703_cast_fp16, y = model_layers_14_self_attn_q_norm_weight_to_fp16)[name = string("x_175_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(533978112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535026752))))[name = string("model_layers_14_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_99_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_k_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_99_cast_fp16")];
            tensor<int32, [4]> var_5712 = const()[name = string("op_5712"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5713_cast_fp16 = reshape(shape = var_5712, x = linear_99_cast_fp16)[name = string("op_5713_cast_fp16")];
            tensor<int32, [4]> x_173_perm_0 = const()[name = string("x_173_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5721_promoted_to_fp16 = const()[name = string("op_5721_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_173_cast_fp16 = transpose(perm = x_173_perm_0, x = var_5713_cast_fp16)[name = string("transpose_68")];
            tensor<fp16, [1, 8, 16, 128]> var_5723_cast_fp16 = pow(x = x_173_cast_fp16, y = var_5721_promoted_to_fp16)[name = string("op_5723_cast_fp16")];
            tensor<int32, [1]> var_5725_axes_0 = const()[name = string("op_5725_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5725_keep_dims_0 = const()[name = string("op_5725_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_5725_cast_fp16 = reduce_mean(axes = var_5725_axes_0, keep_dims = var_5725_keep_dims_0, x = var_5723_cast_fp16)[name = string("op_5725_cast_fp16")];
            fp16 var_5726_to_fp16 = const()[name = string("op_5726_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_5727_cast_fp16 = add(x = var_5725_cast_fp16, y = var_5726_to_fp16)[name = string("op_5727_cast_fp16")];
            fp32 var_5728_epsilon_0 = const()[name = string("op_5728_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_5728_cast_fp16 = rsqrt(epsilon = var_5728_epsilon_0, x = var_5727_cast_fp16)[name = string("op_5728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_5729_cast_fp16 = mul(x = x_173_cast_fp16, y = var_5728_cast_fp16)[name = string("op_5729_cast_fp16")];
            tensor<fp16, [128]> model_layers_14_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_14_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535028864)))];
            tensor<fp16, [1, 8, 16, 128]> x_177_cast_fp16 = mul(x = var_5729_cast_fp16, y = model_layers_14_self_attn_k_norm_weight_to_fp16)[name = string("x_177_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_14_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(535029184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536077824))))[name = string("model_layers_14_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_100_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_v_proj_weight_to_fp16_quantized, x = input_141_cast_fp16)[name = string("linear_100_cast_fp16")];
            tensor<int32, [4]> var_5738 = const()[name = string("op_5738"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_5739_cast_fp16 = reshape(shape = var_5738, x = linear_100_cast_fp16)[name = string("op_5739_cast_fp16")];
            tensor<int32, [4]> v_29_perm_0 = const()[name = string("v_29_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_57_begin_0 = const()[name = string("first_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_57_end_0 = const()[name = string("first_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_57_end_mask_0 = const()[name = string("first_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_57_cast_fp16 = slice_by_index(begin = first_57_begin_0, end = first_57_end_0, end_mask = first_57_end_mask_0, x = x_175_cast_fp16)[name = string("first_57_cast_fp16")];
            tensor<int32, [4]> second_57_begin_0 = const()[name = string("second_57_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_57_end_0 = const()[name = string("second_57_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_57_end_mask_0 = const()[name = string("second_57_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_57_cast_fp16 = slice_by_index(begin = second_57_begin_0, end = second_57_end_0, end_mask = second_57_end_mask_0, x = x_175_cast_fp16)[name = string("second_57_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5753_cast_fp16 = mul(x = first_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5753_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5754_cast_fp16 = mul(x = second_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5754_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5756_cast_fp16 = sub(x = var_5753_cast_fp16, y = var_5754_cast_fp16)[name = string("op_5756_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5757_cast_fp16 = mul(x = second_57_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5757_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5758_cast_fp16 = mul(x = first_57_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5758_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_5760_cast_fp16 = add(x = var_5757_cast_fp16, y = var_5758_cast_fp16)[name = string("op_5760_cast_fp16")];
            int32 var_5762 = const()[name = string("op_5762"), val = int32(-1)];
            bool q_29_interleave_0 = const()[name = string("q_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_29_cast_fp16 = concat(axis = var_5762, interleave = q_29_interleave_0, values = (var_5756_cast_fp16, var_5760_cast_fp16))[name = string("q_29_cast_fp16")];
            tensor<int32, [4]> first_59_begin_0 = const()[name = string("first_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_59_end_0 = const()[name = string("first_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_59_end_mask_0 = const()[name = string("first_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_59_cast_fp16 = slice_by_index(begin = first_59_begin_0, end = first_59_end_0, end_mask = first_59_end_mask_0, x = x_177_cast_fp16)[name = string("first_59_cast_fp16")];
            tensor<int32, [4]> second_59_begin_0 = const()[name = string("second_59_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_59_end_0 = const()[name = string("second_59_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_59_end_mask_0 = const()[name = string("second_59_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_59_cast_fp16 = slice_by_index(begin = second_59_begin_0, end = second_59_end_0, end_mask = second_59_end_mask_0, x = x_177_cast_fp16)[name = string("second_59_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5774_cast_fp16 = mul(x = first_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5774_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5775_cast_fp16 = mul(x = second_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5775_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5777_cast_fp16 = sub(x = var_5774_cast_fp16, y = var_5775_cast_fp16)[name = string("op_5777_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5778_cast_fp16 = mul(x = second_59_cast_fp16, y = cos_1_cast_fp16)[name = string("op_5778_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5779_cast_fp16 = mul(x = first_59_cast_fp16, y = sin_1_cast_fp16)[name = string("op_5779_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_5781_cast_fp16 = add(x = var_5778_cast_fp16, y = var_5779_cast_fp16)[name = string("op_5781_cast_fp16")];
            int32 var_5783 = const()[name = string("op_5783"), val = int32(-1)];
            bool k_29_interleave_0 = const()[name = string("k_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_29_cast_fp16 = concat(axis = var_5783, interleave = k_29_interleave_0, values = (var_5777_cast_fp16, var_5781_cast_fp16))[name = string("k_29_cast_fp16")];
            tensor<int32, [4]> var_5794_begin_0 = const()[name = string("op_5794_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5794_end_0 = const()[name = string("op_5794_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5794_end_mask_0 = const()[name = string("op_5794_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5794_stride_0 = const()[name = string("op_5794_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5794_cast_fp16 = slice_by_index(begin = var_5794_begin_0, end = var_5794_end_0, end_mask = var_5794_end_mask_0, stride = var_5794_stride_0, x = q_29_cast_fp16)[name = string("op_5794_cast_fp16")];
            tensor<int32, [4]> var_5804_begin_0 = const()[name = string("op_5804_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_5804_end_0 = const()[name = string("op_5804_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_5804_end_mask_0 = const()[name = string("op_5804_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_5804_stride_0 = const()[name = string("op_5804_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_5804_cast_fp16 = slice_by_index(begin = var_5804_begin_0, end = var_5804_end_0, end_mask = var_5804_end_mask_0, stride = var_5804_stride_0, x = q_29_cast_fp16)[name = string("op_5804_cast_fp16")];
            int32 var_5806 = const()[name = string("op_5806"), val = int32(2)];
            bool stacked_29_interleave_0 = const()[name = string("stacked_29_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_29_cast_fp16 = concat(axis = var_5806, interleave = stacked_29_interleave_0, values = (var_5794_cast_fp16, var_5804_cast_fp16))[name = string("stacked_29_cast_fp16")];
            tensor<int32, [4]> var_5812_begin_0 = const()[name = string("op_5812_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5812_end_0 = const()[name = string("op_5812_end_0"), val = tensor<int32, [4]>([15, 8, 528, 128])];
            tensor<bool, [4]> var_5812_end_mask_0 = const()[name = string("op_5812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5812_cast_fp16 = slice_by_index(begin = var_5812_begin_0, end = var_5812_end_0, end_mask = var_5812_end_mask_0, x = read_state_0)[name = string("op_5812_cast_fp16")];
            tensor<int32, [4]> var_5822_begin_0 = const()[name = string("op_5822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5822_end_0 = const()[name = string("op_5822_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5822_end_mask_0 = const()[name = string("op_5822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5822_cast_fp16 = slice_by_index(begin = var_5822_begin_0, end = var_5822_end_0, end_mask = var_5822_end_mask_0, x = var_5812_cast_fp16)[name = string("op_5822_cast_fp16")];
            bool var_5831_transpose_x_1 = const()[name = string("op_5831_transpose_x_1"), val = bool(false)];
            bool var_5831_transpose_y_1 = const()[name = string("op_5831_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_5831_cast_fp16 = matmul(transpose_x = var_5831_transpose_x_1, transpose_y = var_5831_transpose_y_1, x = var_5822_cast_fp16, y = stacked_29_cast_fp16)[name = string("op_5831_cast_fp16")];
            tensor<int32, [4]> var_5834_perm_0 = const()[name = string("op_5834_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_5835_to_fp16 = const()[name = string("op_5835_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_5834_cast_fp16 = transpose(perm = var_5834_perm_0, x = var_5831_cast_fp16)[name = string("transpose_67")];
            tensor<fp16, [1, 8, 32, 512]> var_5836_cast_fp16 = mul(x = var_5834_cast_fp16, y = var_5835_to_fp16)[name = string("op_5836_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_29_cast_fp16 = add(x = var_5836_cast_fp16, y = cacheMask)[name = string("cache_weights_29_cast_fp16")];
            bool var_5842_transpose_x_1 = const()[name = string("op_5842_transpose_x_1"), val = bool(false)];
            bool var_5842_transpose_y_1 = const()[name = string("op_5842_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_5842_cast_fp16 = matmul(transpose_x = var_5842_transpose_x_1, transpose_y = var_5842_transpose_y_1, x = stacked_29_cast_fp16, y = k_29_cast_fp16)[name = string("op_5842_cast_fp16")];
            fp16 var_5843_to_fp16 = const()[name = string("op_5843_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_5844_cast_fp16 = mul(x = var_5842_cast_fp16, y = var_5843_to_fp16)[name = string("op_5844_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_29_cast_fp16 = add(x = var_5844_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_29_cast_fp16")];
            int32 var_5848 = const()[name = string("op_5848"), val = int32(-1)];
            bool var_5849_interleave_0 = const()[name = string("op_5849_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_5849_cast_fp16 = concat(axis = var_5848, interleave = var_5849_interleave_0, values = (cache_weights_29_cast_fp16, call_weights_29_cast_fp16))[name = string("op_5849_cast_fp16")];
            int32 var_5850 = const()[name = string("op_5850"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_29_cast_fp16 = softmax(axis = var_5850, x = var_5849_cast_fp16)[name = string("shares_29_cast_fp16")];
            tensor<int32, [4]> var_5857_begin_0 = const()[name = string("op_5857_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5857_end_0 = const()[name = string("op_5857_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_5857_end_mask_0 = const()[name = string("op_5857_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_5857_cast_fp16 = slice_by_index(begin = var_5857_begin_0, end = var_5857_end_0, end_mask = var_5857_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5857_cast_fp16")];
            tensor<int32, [4]> var_5862_begin_0 = const()[name = string("op_5862_begin_0"), val = tensor<int32, [4]>([14, 0, 0, 0])];
            tensor<int32, [4]> var_5862_end_0 = const()[name = string("op_5862_end_0"), val = tensor<int32, [4]>([15, 8, 528, 128])];
            tensor<bool, [4]> var_5862_end_mask_0 = const()[name = string("op_5862_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_5862_cast_fp16 = slice_by_index(begin = var_5862_begin_0, end = var_5862_end_0, end_mask = var_5862_end_mask_0, x = read_state_1)[name = string("op_5862_cast_fp16")];
            tensor<int32, [4]> var_5872_begin_0 = const()[name = string("op_5872_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5872_end_0 = const()[name = string("op_5872_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_5872_end_mask_0 = const()[name = string("op_5872_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_5872_cast_fp16 = slice_by_index(begin = var_5872_begin_0, end = var_5872_end_0, end_mask = var_5872_end_mask_0, x = var_5862_cast_fp16)[name = string("op_5872_cast_fp16")];
            bool var_5878_transpose_x_0 = const()[name = string("op_5878_transpose_x_0"), val = bool(false)];
            bool var_5878_transpose_y_0 = const()[name = string("op_5878_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_5878_cast_fp16 = matmul(transpose_x = var_5878_transpose_x_0, transpose_y = var_5878_transpose_y_0, x = var_5857_cast_fp16, y = var_5872_cast_fp16)[name = string("op_5878_cast_fp16")];
            tensor<int32, [4]> var_5883_begin_0 = const()[name = string("op_5883_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_5883_end_0 = const()[name = string("op_5883_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_5883_end_mask_0 = const()[name = string("op_5883_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_5883_cast_fp16 = slice_by_index(begin = var_5883_begin_0, end = var_5883_end_0, end_mask = var_5883_end_mask_0, x = shares_29_cast_fp16)[name = string("op_5883_cast_fp16")];
            bool var_5884_transpose_x_0 = const()[name = string("op_5884_transpose_x_0"), val = bool(false)];
            bool var_5884_transpose_y_0 = const()[name = string("op_5884_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_29_cast_fp16 = transpose(perm = v_29_perm_0, x = var_5739_cast_fp16)[name = string("transpose_66")];
            tensor<fp16, [1, 8, 32, 128]> var_5884_cast_fp16 = matmul(transpose_x = var_5884_transpose_x_0, transpose_y = var_5884_transpose_y_0, x = var_5883_cast_fp16, y = v_29_cast_fp16)[name = string("op_5884_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_57_cast_fp16 = add(x = var_5878_cast_fp16, y = var_5884_cast_fp16)[name = string("mixed_57_cast_fp16")];
            tensor<int32, [4]> var_5901_begin_0 = const()[name = string("op_5901_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_5901_end_0 = const()[name = string("op_5901_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_5901_end_mask_0 = const()[name = string("op_5901_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5901_cast_fp16 = slice_by_index(begin = var_5901_begin_0, end = var_5901_end_0, end_mask = var_5901_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5901_cast_fp16")];
            tensor<int32, [4]> var_5916_begin_0 = const()[name = string("op_5916_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_5916_end_0 = const()[name = string("op_5916_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_5916_end_mask_0 = const()[name = string("op_5916_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_5916_cast_fp16 = slice_by_index(begin = var_5916_begin_0, end = var_5916_end_0, end_mask = var_5916_end_mask_0, x = mixed_57_cast_fp16)[name = string("op_5916_cast_fp16")];
            int32 var_5918 = const()[name = string("op_5918"), val = int32(1)];
            bool mixed_59_interleave_0 = const()[name = string("mixed_59_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_59_cast_fp16 = concat(axis = var_5918, interleave = mixed_59_interleave_0, values = (var_5901_cast_fp16, var_5916_cast_fp16))[name = string("mixed_59_cast_fp16")];
            tensor<int32, [4]> var_5922_perm_0 = const()[name = string("op_5922_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_5926 = const()[name = string("op_5926"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_5922_cast_fp16 = transpose(perm = var_5922_perm_0, x = mixed_59_cast_fp16)[name = string("transpose_65")];
            tensor<fp16, [1, 16, 2048]> input_143_cast_fp16 = reshape(shape = var_5926, x = var_5922_cast_fp16)[name = string("input_143_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_14_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(536079936))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538177152))))[name = string("model_layers_14_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_101_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_self_attn_o_proj_weight_to_fp16_quantized, x = input_143_cast_fp16)[name = string("linear_101_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_179_cast_fp16 = add(x = x_169_cast_fp16, y = linear_101_cast_fp16)[name = string("x_179_cast_fp16")];
            fp16 var_5937_promoted_to_fp16 = const()[name = string("op_5937_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5939_cast_fp16 = pow(x = x_179_cast_fp16, y = var_5937_promoted_to_fp16)[name = string("op_5939_cast_fp16")];
            tensor<int32, [1]> var_5941_axes_0 = const()[name = string("op_5941_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5941_keep_dims_0 = const()[name = string("op_5941_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5941_cast_fp16 = reduce_mean(axes = var_5941_axes_0, keep_dims = var_5941_keep_dims_0, x = var_5939_cast_fp16)[name = string("op_5941_cast_fp16")];
            fp16 var_5942_to_fp16 = const()[name = string("op_5942_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5943_cast_fp16 = add(x = var_5941_cast_fp16, y = var_5942_to_fp16)[name = string("op_5943_cast_fp16")];
            fp32 var_5944_epsilon_0 = const()[name = string("op_5944_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5944_cast_fp16 = rsqrt(epsilon = var_5944_epsilon_0, x = var_5943_cast_fp16)[name = string("op_5944_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5945_cast_fp16 = mul(x = x_179_cast_fp16, y = var_5944_cast_fp16)[name = string("op_5945_cast_fp16")];
            tensor<fp16, [1024]> model_layers_14_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_14_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538179264)))];
            tensor<fp16, [1, 16, 1024]> input_145_cast_fp16 = mul(x = var_5945_cast_fp16, y = model_layers_14_post_attention_layernorm_weight_to_fp16)[name = string("input_145_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(538181376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541327168))))[name = string("model_layers_14_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_102_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_gate_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_102_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_5950_cast_fp16 = silu(x = linear_102_cast_fp16)[name = string("op_5950_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_14_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(541333376))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544479168))))[name = string("model_layers_14_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_103_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_14_mlp_up_proj_weight_to_fp16_quantized, x = input_145_cast_fp16)[name = string("linear_103_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_149_cast_fp16 = mul(x = var_5950_cast_fp16, y = linear_103_cast_fp16)[name = string("input_149_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_14_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(544485376))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547631168))))[name = string("model_layers_14_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_104_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_14_mlp_down_proj_weight_to_fp16_quantized, x = input_149_cast_fp16)[name = string("linear_104_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_181_cast_fp16 = add(x = x_179_cast_fp16, y = linear_104_cast_fp16)[name = string("x_181_cast_fp16")];
            fp16 var_5964_promoted_to_fp16 = const()[name = string("op_5964_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_5966_cast_fp16 = pow(x = x_181_cast_fp16, y = var_5964_promoted_to_fp16)[name = string("op_5966_cast_fp16")];
            tensor<int32, [1]> var_5968_axes_0 = const()[name = string("op_5968_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5968_keep_dims_0 = const()[name = string("op_5968_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_5968_cast_fp16 = reduce_mean(axes = var_5968_axes_0, keep_dims = var_5968_keep_dims_0, x = var_5966_cast_fp16)[name = string("op_5968_cast_fp16")];
            fp16 var_5969_to_fp16 = const()[name = string("op_5969_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_5970_cast_fp16 = add(x = var_5968_cast_fp16, y = var_5969_to_fp16)[name = string("op_5970_cast_fp16")];
            fp32 var_5971_epsilon_0 = const()[name = string("op_5971_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_5971_cast_fp16 = rsqrt(epsilon = var_5971_epsilon_0, x = var_5970_cast_fp16)[name = string("op_5971_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_5972_cast_fp16 = mul(x = x_181_cast_fp16, y = var_5971_cast_fp16)[name = string("op_5972_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547633280)))];
            tensor<fp16, [1, 16, 1024]> input_151_cast_fp16 = mul(x = var_5972_cast_fp16, y = model_layers_15_input_layernorm_weight_to_fp16)[name = string("input_151_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_15_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(547635392))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549732608))))[name = string("model_layers_15_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_105_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_15_self_attn_q_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_105_cast_fp16")];
            tensor<int32, [4]> var_5981 = const()[name = string("op_5981"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_5982_cast_fp16 = reshape(shape = var_5981, x = linear_105_cast_fp16)[name = string("op_5982_cast_fp16")];
            tensor<int32, [4]> x_183_perm_0 = const()[name = string("x_183_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_5990_promoted_to_fp16 = const()[name = string("op_5990_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_183_cast_fp16 = transpose(perm = x_183_perm_0, x = var_5982_cast_fp16)[name = string("transpose_64")];
            tensor<fp16, [1, 16, 16, 128]> var_5992_cast_fp16 = pow(x = x_183_cast_fp16, y = var_5990_promoted_to_fp16)[name = string("op_5992_cast_fp16")];
            tensor<int32, [1]> var_5994_axes_0 = const()[name = string("op_5994_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_5994_keep_dims_0 = const()[name = string("op_5994_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_5994_cast_fp16 = reduce_mean(axes = var_5994_axes_0, keep_dims = var_5994_keep_dims_0, x = var_5992_cast_fp16)[name = string("op_5994_cast_fp16")];
            fp16 var_5995_to_fp16 = const()[name = string("op_5995_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_5996_cast_fp16 = add(x = var_5994_cast_fp16, y = var_5995_to_fp16)[name = string("op_5996_cast_fp16")];
            fp32 var_5997_epsilon_0 = const()[name = string("op_5997_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_5997_cast_fp16 = rsqrt(epsilon = var_5997_epsilon_0, x = var_5996_cast_fp16)[name = string("op_5997_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_5998_cast_fp16 = mul(x = x_183_cast_fp16, y = var_5997_cast_fp16)[name = string("op_5998_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549736768)))];
            tensor<fp16, [1, 16, 16, 128]> x_187_cast_fp16 = mul(x = var_5998_cast_fp16, y = model_layers_15_self_attn_q_norm_weight_to_fp16)[name = string("x_187_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(549737088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550785728))))[name = string("model_layers_15_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_106_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_k_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_106_cast_fp16")];
            tensor<int32, [4]> var_6007 = const()[name = string("op_6007"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6008_cast_fp16 = reshape(shape = var_6007, x = linear_106_cast_fp16)[name = string("op_6008_cast_fp16")];
            tensor<int32, [4]> x_185_perm_0 = const()[name = string("x_185_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6016_promoted_to_fp16 = const()[name = string("op_6016_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_185_cast_fp16 = transpose(perm = x_185_perm_0, x = var_6008_cast_fp16)[name = string("transpose_63")];
            tensor<fp16, [1, 8, 16, 128]> var_6018_cast_fp16 = pow(x = x_185_cast_fp16, y = var_6016_promoted_to_fp16)[name = string("op_6018_cast_fp16")];
            tensor<int32, [1]> var_6020_axes_0 = const()[name = string("op_6020_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6020_keep_dims_0 = const()[name = string("op_6020_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6020_cast_fp16 = reduce_mean(axes = var_6020_axes_0, keep_dims = var_6020_keep_dims_0, x = var_6018_cast_fp16)[name = string("op_6020_cast_fp16")];
            fp16 var_6021_to_fp16 = const()[name = string("op_6021_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6022_cast_fp16 = add(x = var_6020_cast_fp16, y = var_6021_to_fp16)[name = string("op_6022_cast_fp16")];
            fp32 var_6023_epsilon_0 = const()[name = string("op_6023_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6023_cast_fp16 = rsqrt(epsilon = var_6023_epsilon_0, x = var_6022_cast_fp16)[name = string("op_6023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6024_cast_fp16 = mul(x = x_185_cast_fp16, y = var_6023_cast_fp16)[name = string("op_6024_cast_fp16")];
            tensor<fp16, [128]> model_layers_15_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_15_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550787840)))];
            tensor<fp16, [1, 8, 16, 128]> x_189_cast_fp16 = mul(x = var_6024_cast_fp16, y = model_layers_15_self_attn_k_norm_weight_to_fp16)[name = string("x_189_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_15_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(550788160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551836800))))[name = string("model_layers_15_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_107_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_v_proj_weight_to_fp16_quantized, x = input_151_cast_fp16)[name = string("linear_107_cast_fp16")];
            tensor<int32, [4]> var_6033 = const()[name = string("op_6033"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6034_cast_fp16 = reshape(shape = var_6033, x = linear_107_cast_fp16)[name = string("op_6034_cast_fp16")];
            tensor<int32, [4]> v_31_perm_0 = const()[name = string("v_31_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_61_begin_0 = const()[name = string("first_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_61_end_0 = const()[name = string("first_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_61_end_mask_0 = const()[name = string("first_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_61_cast_fp16 = slice_by_index(begin = first_61_begin_0, end = first_61_end_0, end_mask = first_61_end_mask_0, x = x_187_cast_fp16)[name = string("first_61_cast_fp16")];
            tensor<int32, [4]> second_61_begin_0 = const()[name = string("second_61_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_61_end_0 = const()[name = string("second_61_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_61_end_mask_0 = const()[name = string("second_61_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_61_cast_fp16 = slice_by_index(begin = second_61_begin_0, end = second_61_end_0, end_mask = second_61_end_mask_0, x = x_187_cast_fp16)[name = string("second_61_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6048_cast_fp16 = mul(x = first_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6048_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6049_cast_fp16 = mul(x = second_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6049_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6051_cast_fp16 = sub(x = var_6048_cast_fp16, y = var_6049_cast_fp16)[name = string("op_6051_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6052_cast_fp16 = mul(x = second_61_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6052_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6053_cast_fp16 = mul(x = first_61_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6053_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6055_cast_fp16 = add(x = var_6052_cast_fp16, y = var_6053_cast_fp16)[name = string("op_6055_cast_fp16")];
            int32 var_6057 = const()[name = string("op_6057"), val = int32(-1)];
            bool q_31_interleave_0 = const()[name = string("q_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_31_cast_fp16 = concat(axis = var_6057, interleave = q_31_interleave_0, values = (var_6051_cast_fp16, var_6055_cast_fp16))[name = string("q_31_cast_fp16")];
            tensor<int32, [4]> first_63_begin_0 = const()[name = string("first_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_63_end_0 = const()[name = string("first_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_63_end_mask_0 = const()[name = string("first_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_63_cast_fp16 = slice_by_index(begin = first_63_begin_0, end = first_63_end_0, end_mask = first_63_end_mask_0, x = x_189_cast_fp16)[name = string("first_63_cast_fp16")];
            tensor<int32, [4]> second_63_begin_0 = const()[name = string("second_63_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_63_end_0 = const()[name = string("second_63_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_63_end_mask_0 = const()[name = string("second_63_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_63_cast_fp16 = slice_by_index(begin = second_63_begin_0, end = second_63_end_0, end_mask = second_63_end_mask_0, x = x_189_cast_fp16)[name = string("second_63_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6069_cast_fp16 = mul(x = first_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6069_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6070_cast_fp16 = mul(x = second_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6070_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6072_cast_fp16 = sub(x = var_6069_cast_fp16, y = var_6070_cast_fp16)[name = string("op_6072_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6073_cast_fp16 = mul(x = second_63_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6073_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6074_cast_fp16 = mul(x = first_63_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6074_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6076_cast_fp16 = add(x = var_6073_cast_fp16, y = var_6074_cast_fp16)[name = string("op_6076_cast_fp16")];
            int32 var_6078 = const()[name = string("op_6078"), val = int32(-1)];
            bool k_31_interleave_0 = const()[name = string("k_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_31_cast_fp16 = concat(axis = var_6078, interleave = k_31_interleave_0, values = (var_6072_cast_fp16, var_6076_cast_fp16))[name = string("k_31_cast_fp16")];
            tensor<int32, [4]> var_6089_begin_0 = const()[name = string("op_6089_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6089_end_0 = const()[name = string("op_6089_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6089_end_mask_0 = const()[name = string("op_6089_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6089_stride_0 = const()[name = string("op_6089_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6089_cast_fp16 = slice_by_index(begin = var_6089_begin_0, end = var_6089_end_0, end_mask = var_6089_end_mask_0, stride = var_6089_stride_0, x = q_31_cast_fp16)[name = string("op_6089_cast_fp16")];
            tensor<int32, [4]> var_6099_begin_0 = const()[name = string("op_6099_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6099_end_0 = const()[name = string("op_6099_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6099_end_mask_0 = const()[name = string("op_6099_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6099_stride_0 = const()[name = string("op_6099_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6099_cast_fp16 = slice_by_index(begin = var_6099_begin_0, end = var_6099_end_0, end_mask = var_6099_end_mask_0, stride = var_6099_stride_0, x = q_31_cast_fp16)[name = string("op_6099_cast_fp16")];
            int32 var_6101 = const()[name = string("op_6101"), val = int32(2)];
            bool stacked_31_interleave_0 = const()[name = string("stacked_31_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_31_cast_fp16 = concat(axis = var_6101, interleave = stacked_31_interleave_0, values = (var_6089_cast_fp16, var_6099_cast_fp16))[name = string("stacked_31_cast_fp16")];
            tensor<int32, [4]> var_6107_begin_0 = const()[name = string("op_6107_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6107_end_0 = const()[name = string("op_6107_end_0"), val = tensor<int32, [4]>([16, 8, 528, 128])];
            tensor<bool, [4]> var_6107_end_mask_0 = const()[name = string("op_6107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6107_cast_fp16 = slice_by_index(begin = var_6107_begin_0, end = var_6107_end_0, end_mask = var_6107_end_mask_0, x = read_state_0)[name = string("op_6107_cast_fp16")];
            tensor<int32, [4]> var_6117_begin_0 = const()[name = string("op_6117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6117_end_0 = const()[name = string("op_6117_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6117_end_mask_0 = const()[name = string("op_6117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6117_cast_fp16 = slice_by_index(begin = var_6117_begin_0, end = var_6117_end_0, end_mask = var_6117_end_mask_0, x = var_6107_cast_fp16)[name = string("op_6117_cast_fp16")];
            bool var_6126_transpose_x_1 = const()[name = string("op_6126_transpose_x_1"), val = bool(false)];
            bool var_6126_transpose_y_1 = const()[name = string("op_6126_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6126_cast_fp16 = matmul(transpose_x = var_6126_transpose_x_1, transpose_y = var_6126_transpose_y_1, x = var_6117_cast_fp16, y = stacked_31_cast_fp16)[name = string("op_6126_cast_fp16")];
            tensor<int32, [4]> var_6129_perm_0 = const()[name = string("op_6129_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6130_to_fp16 = const()[name = string("op_6130_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6129_cast_fp16 = transpose(perm = var_6129_perm_0, x = var_6126_cast_fp16)[name = string("transpose_62")];
            tensor<fp16, [1, 8, 32, 512]> var_6131_cast_fp16 = mul(x = var_6129_cast_fp16, y = var_6130_to_fp16)[name = string("op_6131_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_31_cast_fp16 = add(x = var_6131_cast_fp16, y = cacheMask)[name = string("cache_weights_31_cast_fp16")];
            bool var_6137_transpose_x_1 = const()[name = string("op_6137_transpose_x_1"), val = bool(false)];
            bool var_6137_transpose_y_1 = const()[name = string("op_6137_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6137_cast_fp16 = matmul(transpose_x = var_6137_transpose_x_1, transpose_y = var_6137_transpose_y_1, x = stacked_31_cast_fp16, y = k_31_cast_fp16)[name = string("op_6137_cast_fp16")];
            fp16 var_6138_to_fp16 = const()[name = string("op_6138_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6139_cast_fp16 = mul(x = var_6137_cast_fp16, y = var_6138_to_fp16)[name = string("op_6139_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_31_cast_fp16 = add(x = var_6139_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_31_cast_fp16")];
            int32 var_6143 = const()[name = string("op_6143"), val = int32(-1)];
            bool var_6144_interleave_0 = const()[name = string("op_6144_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6144_cast_fp16 = concat(axis = var_6143, interleave = var_6144_interleave_0, values = (cache_weights_31_cast_fp16, call_weights_31_cast_fp16))[name = string("op_6144_cast_fp16")];
            int32 var_6145 = const()[name = string("op_6145"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_31_cast_fp16 = softmax(axis = var_6145, x = var_6144_cast_fp16)[name = string("shares_31_cast_fp16")];
            tensor<int32, [4]> var_6152_begin_0 = const()[name = string("op_6152_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6152_end_0 = const()[name = string("op_6152_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6152_end_mask_0 = const()[name = string("op_6152_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6152_cast_fp16 = slice_by_index(begin = var_6152_begin_0, end = var_6152_end_0, end_mask = var_6152_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6152_cast_fp16")];
            tensor<int32, [4]> var_6157_begin_0 = const()[name = string("op_6157_begin_0"), val = tensor<int32, [4]>([15, 0, 0, 0])];
            tensor<int32, [4]> var_6157_end_0 = const()[name = string("op_6157_end_0"), val = tensor<int32, [4]>([16, 8, 528, 128])];
            tensor<bool, [4]> var_6157_end_mask_0 = const()[name = string("op_6157_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6157_cast_fp16 = slice_by_index(begin = var_6157_begin_0, end = var_6157_end_0, end_mask = var_6157_end_mask_0, x = read_state_1)[name = string("op_6157_cast_fp16")];
            tensor<int32, [4]> var_6167_begin_0 = const()[name = string("op_6167_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6167_end_0 = const()[name = string("op_6167_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6167_end_mask_0 = const()[name = string("op_6167_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6167_cast_fp16 = slice_by_index(begin = var_6167_begin_0, end = var_6167_end_0, end_mask = var_6167_end_mask_0, x = var_6157_cast_fp16)[name = string("op_6167_cast_fp16")];
            bool var_6173_transpose_x_0 = const()[name = string("op_6173_transpose_x_0"), val = bool(false)];
            bool var_6173_transpose_y_0 = const()[name = string("op_6173_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6173_cast_fp16 = matmul(transpose_x = var_6173_transpose_x_0, transpose_y = var_6173_transpose_y_0, x = var_6152_cast_fp16, y = var_6167_cast_fp16)[name = string("op_6173_cast_fp16")];
            tensor<int32, [4]> var_6178_begin_0 = const()[name = string("op_6178_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6178_end_0 = const()[name = string("op_6178_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6178_end_mask_0 = const()[name = string("op_6178_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6178_cast_fp16 = slice_by_index(begin = var_6178_begin_0, end = var_6178_end_0, end_mask = var_6178_end_mask_0, x = shares_31_cast_fp16)[name = string("op_6178_cast_fp16")];
            bool var_6179_transpose_x_0 = const()[name = string("op_6179_transpose_x_0"), val = bool(false)];
            bool var_6179_transpose_y_0 = const()[name = string("op_6179_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_31_cast_fp16 = transpose(perm = v_31_perm_0, x = var_6034_cast_fp16)[name = string("transpose_61")];
            tensor<fp16, [1, 8, 32, 128]> var_6179_cast_fp16 = matmul(transpose_x = var_6179_transpose_x_0, transpose_y = var_6179_transpose_y_0, x = var_6178_cast_fp16, y = v_31_cast_fp16)[name = string("op_6179_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_61_cast_fp16 = add(x = var_6173_cast_fp16, y = var_6179_cast_fp16)[name = string("mixed_61_cast_fp16")];
            tensor<int32, [4]> var_6196_begin_0 = const()[name = string("op_6196_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6196_end_0 = const()[name = string("op_6196_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6196_end_mask_0 = const()[name = string("op_6196_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6196_cast_fp16 = slice_by_index(begin = var_6196_begin_0, end = var_6196_end_0, end_mask = var_6196_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6196_cast_fp16")];
            tensor<int32, [4]> var_6211_begin_0 = const()[name = string("op_6211_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6211_end_0 = const()[name = string("op_6211_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6211_end_mask_0 = const()[name = string("op_6211_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6211_cast_fp16 = slice_by_index(begin = var_6211_begin_0, end = var_6211_end_0, end_mask = var_6211_end_mask_0, x = mixed_61_cast_fp16)[name = string("op_6211_cast_fp16")];
            int32 var_6213 = const()[name = string("op_6213"), val = int32(1)];
            bool mixed_63_interleave_0 = const()[name = string("mixed_63_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_63_cast_fp16 = concat(axis = var_6213, interleave = mixed_63_interleave_0, values = (var_6196_cast_fp16, var_6211_cast_fp16))[name = string("mixed_63_cast_fp16")];
            tensor<int32, [4]> var_6217_perm_0 = const()[name = string("op_6217_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6221 = const()[name = string("op_6221"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6217_cast_fp16 = transpose(perm = var_6217_perm_0, x = mixed_63_cast_fp16)[name = string("transpose_60")];
            tensor<fp16, [1, 16, 2048]> input_153_cast_fp16 = reshape(shape = var_6221, x = var_6217_cast_fp16)[name = string("input_153_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_15_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(551838912))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553936128))))[name = string("model_layers_15_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_108_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_self_attn_o_proj_weight_to_fp16_quantized, x = input_153_cast_fp16)[name = string("linear_108_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_191_cast_fp16 = add(x = x_181_cast_fp16, y = linear_108_cast_fp16)[name = string("x_191_cast_fp16")];
            fp16 var_6232_promoted_to_fp16 = const()[name = string("op_6232_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6234_cast_fp16 = pow(x = x_191_cast_fp16, y = var_6232_promoted_to_fp16)[name = string("op_6234_cast_fp16")];
            tensor<int32, [1]> var_6236_axes_0 = const()[name = string("op_6236_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6236_keep_dims_0 = const()[name = string("op_6236_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6236_cast_fp16 = reduce_mean(axes = var_6236_axes_0, keep_dims = var_6236_keep_dims_0, x = var_6234_cast_fp16)[name = string("op_6236_cast_fp16")];
            fp16 var_6237_to_fp16 = const()[name = string("op_6237_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6238_cast_fp16 = add(x = var_6236_cast_fp16, y = var_6237_to_fp16)[name = string("op_6238_cast_fp16")];
            fp32 var_6239_epsilon_0 = const()[name = string("op_6239_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6239_cast_fp16 = rsqrt(epsilon = var_6239_epsilon_0, x = var_6238_cast_fp16)[name = string("op_6239_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6240_cast_fp16 = mul(x = x_191_cast_fp16, y = var_6239_cast_fp16)[name = string("op_6240_cast_fp16")];
            tensor<fp16, [1024]> model_layers_15_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_15_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553938240)))];
            tensor<fp16, [1, 16, 1024]> input_155_cast_fp16 = mul(x = var_6240_cast_fp16, y = model_layers_15_post_attention_layernorm_weight_to_fp16)[name = string("input_155_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(553940352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557086144))))[name = string("model_layers_15_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_109_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_gate_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_109_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6245_cast_fp16 = silu(x = linear_109_cast_fp16)[name = string("op_6245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_15_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(557092352))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560238144))))[name = string("model_layers_15_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_110_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_15_mlp_up_proj_weight_to_fp16_quantized, x = input_155_cast_fp16)[name = string("linear_110_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_159_cast_fp16 = mul(x = var_6245_cast_fp16, y = linear_110_cast_fp16)[name = string("input_159_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_15_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(560244352))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563390144))))[name = string("model_layers_15_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_111_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_15_mlp_down_proj_weight_to_fp16_quantized, x = input_159_cast_fp16)[name = string("linear_111_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_193_cast_fp16 = add(x = x_191_cast_fp16, y = linear_111_cast_fp16)[name = string("x_193_cast_fp16")];
            fp16 var_6259_promoted_to_fp16 = const()[name = string("op_6259_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6261_cast_fp16 = pow(x = x_193_cast_fp16, y = var_6259_promoted_to_fp16)[name = string("op_6261_cast_fp16")];
            tensor<int32, [1]> var_6263_axes_0 = const()[name = string("op_6263_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6263_keep_dims_0 = const()[name = string("op_6263_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6263_cast_fp16 = reduce_mean(axes = var_6263_axes_0, keep_dims = var_6263_keep_dims_0, x = var_6261_cast_fp16)[name = string("op_6263_cast_fp16")];
            fp16 var_6264_to_fp16 = const()[name = string("op_6264_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6265_cast_fp16 = add(x = var_6263_cast_fp16, y = var_6264_to_fp16)[name = string("op_6265_cast_fp16")];
            fp32 var_6266_epsilon_0 = const()[name = string("op_6266_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6266_cast_fp16 = rsqrt(epsilon = var_6266_epsilon_0, x = var_6265_cast_fp16)[name = string("op_6266_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6267_cast_fp16 = mul(x = x_193_cast_fp16, y = var_6266_cast_fp16)[name = string("op_6267_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563392256)))];
            tensor<fp16, [1, 16, 1024]> input_161_cast_fp16 = mul(x = var_6267_cast_fp16, y = model_layers_16_input_layernorm_weight_to_fp16)[name = string("input_161_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_16_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(563394368))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565491584))))[name = string("model_layers_16_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_112_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_16_self_attn_q_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_112_cast_fp16")];
            tensor<int32, [4]> var_6276 = const()[name = string("op_6276"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6277_cast_fp16 = reshape(shape = var_6276, x = linear_112_cast_fp16)[name = string("op_6277_cast_fp16")];
            tensor<int32, [4]> x_195_perm_0 = const()[name = string("x_195_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6285_promoted_to_fp16 = const()[name = string("op_6285_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_195_cast_fp16 = transpose(perm = x_195_perm_0, x = var_6277_cast_fp16)[name = string("transpose_59")];
            tensor<fp16, [1, 16, 16, 128]> var_6287_cast_fp16 = pow(x = x_195_cast_fp16, y = var_6285_promoted_to_fp16)[name = string("op_6287_cast_fp16")];
            tensor<int32, [1]> var_6289_axes_0 = const()[name = string("op_6289_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6289_keep_dims_0 = const()[name = string("op_6289_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6289_cast_fp16 = reduce_mean(axes = var_6289_axes_0, keep_dims = var_6289_keep_dims_0, x = var_6287_cast_fp16)[name = string("op_6289_cast_fp16")];
            fp16 var_6290_to_fp16 = const()[name = string("op_6290_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6291_cast_fp16 = add(x = var_6289_cast_fp16, y = var_6290_to_fp16)[name = string("op_6291_cast_fp16")];
            fp32 var_6292_epsilon_0 = const()[name = string("op_6292_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6292_cast_fp16 = rsqrt(epsilon = var_6292_epsilon_0, x = var_6291_cast_fp16)[name = string("op_6292_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6293_cast_fp16 = mul(x = x_195_cast_fp16, y = var_6292_cast_fp16)[name = string("op_6293_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565495744)))];
            tensor<fp16, [1, 16, 16, 128]> x_199_cast_fp16 = mul(x = var_6293_cast_fp16, y = model_layers_16_self_attn_q_norm_weight_to_fp16)[name = string("x_199_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(565496064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566544704))))[name = string("model_layers_16_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_113_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_k_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_113_cast_fp16")];
            tensor<int32, [4]> var_6302 = const()[name = string("op_6302"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6303_cast_fp16 = reshape(shape = var_6302, x = linear_113_cast_fp16)[name = string("op_6303_cast_fp16")];
            tensor<int32, [4]> x_197_perm_0 = const()[name = string("x_197_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6311_promoted_to_fp16 = const()[name = string("op_6311_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_197_cast_fp16 = transpose(perm = x_197_perm_0, x = var_6303_cast_fp16)[name = string("transpose_58")];
            tensor<fp16, [1, 8, 16, 128]> var_6313_cast_fp16 = pow(x = x_197_cast_fp16, y = var_6311_promoted_to_fp16)[name = string("op_6313_cast_fp16")];
            tensor<int32, [1]> var_6315_axes_0 = const()[name = string("op_6315_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6315_keep_dims_0 = const()[name = string("op_6315_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6315_cast_fp16 = reduce_mean(axes = var_6315_axes_0, keep_dims = var_6315_keep_dims_0, x = var_6313_cast_fp16)[name = string("op_6315_cast_fp16")];
            fp16 var_6316_to_fp16 = const()[name = string("op_6316_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6317_cast_fp16 = add(x = var_6315_cast_fp16, y = var_6316_to_fp16)[name = string("op_6317_cast_fp16")];
            fp32 var_6318_epsilon_0 = const()[name = string("op_6318_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6318_cast_fp16 = rsqrt(epsilon = var_6318_epsilon_0, x = var_6317_cast_fp16)[name = string("op_6318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6319_cast_fp16 = mul(x = x_197_cast_fp16, y = var_6318_cast_fp16)[name = string("op_6319_cast_fp16")];
            tensor<fp16, [128]> model_layers_16_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_16_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566546816)))];
            tensor<fp16, [1, 8, 16, 128]> x_201_cast_fp16 = mul(x = var_6319_cast_fp16, y = model_layers_16_self_attn_k_norm_weight_to_fp16)[name = string("x_201_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_16_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(566547136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567595776))))[name = string("model_layers_16_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_114_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_v_proj_weight_to_fp16_quantized, x = input_161_cast_fp16)[name = string("linear_114_cast_fp16")];
            tensor<int32, [4]> var_6328 = const()[name = string("op_6328"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6329_cast_fp16 = reshape(shape = var_6328, x = linear_114_cast_fp16)[name = string("op_6329_cast_fp16")];
            tensor<int32, [4]> v_33_perm_0 = const()[name = string("v_33_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_65_begin_0 = const()[name = string("first_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_65_end_0 = const()[name = string("first_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_65_end_mask_0 = const()[name = string("first_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_65_cast_fp16 = slice_by_index(begin = first_65_begin_0, end = first_65_end_0, end_mask = first_65_end_mask_0, x = x_199_cast_fp16)[name = string("first_65_cast_fp16")];
            tensor<int32, [4]> second_65_begin_0 = const()[name = string("second_65_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_65_end_0 = const()[name = string("second_65_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_65_end_mask_0 = const()[name = string("second_65_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_65_cast_fp16 = slice_by_index(begin = second_65_begin_0, end = second_65_end_0, end_mask = second_65_end_mask_0, x = x_199_cast_fp16)[name = string("second_65_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6343_cast_fp16 = mul(x = first_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6343_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6344_cast_fp16 = mul(x = second_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6344_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6346_cast_fp16 = sub(x = var_6343_cast_fp16, y = var_6344_cast_fp16)[name = string("op_6346_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6347_cast_fp16 = mul(x = second_65_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6347_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6348_cast_fp16 = mul(x = first_65_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6348_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6350_cast_fp16 = add(x = var_6347_cast_fp16, y = var_6348_cast_fp16)[name = string("op_6350_cast_fp16")];
            int32 var_6352 = const()[name = string("op_6352"), val = int32(-1)];
            bool q_33_interleave_0 = const()[name = string("q_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_33_cast_fp16 = concat(axis = var_6352, interleave = q_33_interleave_0, values = (var_6346_cast_fp16, var_6350_cast_fp16))[name = string("q_33_cast_fp16")];
            tensor<int32, [4]> first_67_begin_0 = const()[name = string("first_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_67_end_0 = const()[name = string("first_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_67_end_mask_0 = const()[name = string("first_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_67_cast_fp16 = slice_by_index(begin = first_67_begin_0, end = first_67_end_0, end_mask = first_67_end_mask_0, x = x_201_cast_fp16)[name = string("first_67_cast_fp16")];
            tensor<int32, [4]> second_67_begin_0 = const()[name = string("second_67_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_67_end_0 = const()[name = string("second_67_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_67_end_mask_0 = const()[name = string("second_67_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_67_cast_fp16 = slice_by_index(begin = second_67_begin_0, end = second_67_end_0, end_mask = second_67_end_mask_0, x = x_201_cast_fp16)[name = string("second_67_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6364_cast_fp16 = mul(x = first_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6364_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6365_cast_fp16 = mul(x = second_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6365_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6367_cast_fp16 = sub(x = var_6364_cast_fp16, y = var_6365_cast_fp16)[name = string("op_6367_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6368_cast_fp16 = mul(x = second_67_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6368_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6369_cast_fp16 = mul(x = first_67_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6369_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6371_cast_fp16 = add(x = var_6368_cast_fp16, y = var_6369_cast_fp16)[name = string("op_6371_cast_fp16")];
            int32 var_6373 = const()[name = string("op_6373"), val = int32(-1)];
            bool k_33_interleave_0 = const()[name = string("k_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_33_cast_fp16 = concat(axis = var_6373, interleave = k_33_interleave_0, values = (var_6367_cast_fp16, var_6371_cast_fp16))[name = string("k_33_cast_fp16")];
            tensor<int32, [4]> var_6384_begin_0 = const()[name = string("op_6384_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6384_end_0 = const()[name = string("op_6384_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6384_end_mask_0 = const()[name = string("op_6384_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6384_stride_0 = const()[name = string("op_6384_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6384_cast_fp16 = slice_by_index(begin = var_6384_begin_0, end = var_6384_end_0, end_mask = var_6384_end_mask_0, stride = var_6384_stride_0, x = q_33_cast_fp16)[name = string("op_6384_cast_fp16")];
            tensor<int32, [4]> var_6394_begin_0 = const()[name = string("op_6394_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6394_end_0 = const()[name = string("op_6394_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6394_end_mask_0 = const()[name = string("op_6394_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6394_stride_0 = const()[name = string("op_6394_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6394_cast_fp16 = slice_by_index(begin = var_6394_begin_0, end = var_6394_end_0, end_mask = var_6394_end_mask_0, stride = var_6394_stride_0, x = q_33_cast_fp16)[name = string("op_6394_cast_fp16")];
            int32 var_6396 = const()[name = string("op_6396"), val = int32(2)];
            bool stacked_33_interleave_0 = const()[name = string("stacked_33_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_33_cast_fp16 = concat(axis = var_6396, interleave = stacked_33_interleave_0, values = (var_6384_cast_fp16, var_6394_cast_fp16))[name = string("stacked_33_cast_fp16")];
            tensor<int32, [4]> var_6402_begin_0 = const()[name = string("op_6402_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6402_end_0 = const()[name = string("op_6402_end_0"), val = tensor<int32, [4]>([17, 8, 528, 128])];
            tensor<bool, [4]> var_6402_end_mask_0 = const()[name = string("op_6402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6402_cast_fp16 = slice_by_index(begin = var_6402_begin_0, end = var_6402_end_0, end_mask = var_6402_end_mask_0, x = read_state_0)[name = string("op_6402_cast_fp16")];
            tensor<int32, [4]> var_6412_begin_0 = const()[name = string("op_6412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6412_end_0 = const()[name = string("op_6412_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6412_end_mask_0 = const()[name = string("op_6412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6412_cast_fp16 = slice_by_index(begin = var_6412_begin_0, end = var_6412_end_0, end_mask = var_6412_end_mask_0, x = var_6402_cast_fp16)[name = string("op_6412_cast_fp16")];
            bool var_6421_transpose_x_1 = const()[name = string("op_6421_transpose_x_1"), val = bool(false)];
            bool var_6421_transpose_y_1 = const()[name = string("op_6421_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6421_cast_fp16 = matmul(transpose_x = var_6421_transpose_x_1, transpose_y = var_6421_transpose_y_1, x = var_6412_cast_fp16, y = stacked_33_cast_fp16)[name = string("op_6421_cast_fp16")];
            tensor<int32, [4]> var_6424_perm_0 = const()[name = string("op_6424_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6425_to_fp16 = const()[name = string("op_6425_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6424_cast_fp16 = transpose(perm = var_6424_perm_0, x = var_6421_cast_fp16)[name = string("transpose_57")];
            tensor<fp16, [1, 8, 32, 512]> var_6426_cast_fp16 = mul(x = var_6424_cast_fp16, y = var_6425_to_fp16)[name = string("op_6426_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_33_cast_fp16 = add(x = var_6426_cast_fp16, y = cacheMask)[name = string("cache_weights_33_cast_fp16")];
            bool var_6432_transpose_x_1 = const()[name = string("op_6432_transpose_x_1"), val = bool(false)];
            bool var_6432_transpose_y_1 = const()[name = string("op_6432_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6432_cast_fp16 = matmul(transpose_x = var_6432_transpose_x_1, transpose_y = var_6432_transpose_y_1, x = stacked_33_cast_fp16, y = k_33_cast_fp16)[name = string("op_6432_cast_fp16")];
            fp16 var_6433_to_fp16 = const()[name = string("op_6433_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6434_cast_fp16 = mul(x = var_6432_cast_fp16, y = var_6433_to_fp16)[name = string("op_6434_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_33_cast_fp16 = add(x = var_6434_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_33_cast_fp16")];
            int32 var_6438 = const()[name = string("op_6438"), val = int32(-1)];
            bool var_6439_interleave_0 = const()[name = string("op_6439_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6439_cast_fp16 = concat(axis = var_6438, interleave = var_6439_interleave_0, values = (cache_weights_33_cast_fp16, call_weights_33_cast_fp16))[name = string("op_6439_cast_fp16")];
            int32 var_6440 = const()[name = string("op_6440"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_33_cast_fp16 = softmax(axis = var_6440, x = var_6439_cast_fp16)[name = string("shares_33_cast_fp16")];
            tensor<int32, [4]> var_6447_begin_0 = const()[name = string("op_6447_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6447_end_0 = const()[name = string("op_6447_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6447_end_mask_0 = const()[name = string("op_6447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6447_cast_fp16 = slice_by_index(begin = var_6447_begin_0, end = var_6447_end_0, end_mask = var_6447_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6447_cast_fp16")];
            tensor<int32, [4]> var_6452_begin_0 = const()[name = string("op_6452_begin_0"), val = tensor<int32, [4]>([16, 0, 0, 0])];
            tensor<int32, [4]> var_6452_end_0 = const()[name = string("op_6452_end_0"), val = tensor<int32, [4]>([17, 8, 528, 128])];
            tensor<bool, [4]> var_6452_end_mask_0 = const()[name = string("op_6452_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6452_cast_fp16 = slice_by_index(begin = var_6452_begin_0, end = var_6452_end_0, end_mask = var_6452_end_mask_0, x = read_state_1)[name = string("op_6452_cast_fp16")];
            tensor<int32, [4]> var_6462_begin_0 = const()[name = string("op_6462_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6462_end_0 = const()[name = string("op_6462_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6462_end_mask_0 = const()[name = string("op_6462_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6462_cast_fp16 = slice_by_index(begin = var_6462_begin_0, end = var_6462_end_0, end_mask = var_6462_end_mask_0, x = var_6452_cast_fp16)[name = string("op_6462_cast_fp16")];
            bool var_6468_transpose_x_0 = const()[name = string("op_6468_transpose_x_0"), val = bool(false)];
            bool var_6468_transpose_y_0 = const()[name = string("op_6468_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6468_cast_fp16 = matmul(transpose_x = var_6468_transpose_x_0, transpose_y = var_6468_transpose_y_0, x = var_6447_cast_fp16, y = var_6462_cast_fp16)[name = string("op_6468_cast_fp16")];
            tensor<int32, [4]> var_6473_begin_0 = const()[name = string("op_6473_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6473_end_0 = const()[name = string("op_6473_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6473_end_mask_0 = const()[name = string("op_6473_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6473_cast_fp16 = slice_by_index(begin = var_6473_begin_0, end = var_6473_end_0, end_mask = var_6473_end_mask_0, x = shares_33_cast_fp16)[name = string("op_6473_cast_fp16")];
            bool var_6474_transpose_x_0 = const()[name = string("op_6474_transpose_x_0"), val = bool(false)];
            bool var_6474_transpose_y_0 = const()[name = string("op_6474_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_33_cast_fp16 = transpose(perm = v_33_perm_0, x = var_6329_cast_fp16)[name = string("transpose_56")];
            tensor<fp16, [1, 8, 32, 128]> var_6474_cast_fp16 = matmul(transpose_x = var_6474_transpose_x_0, transpose_y = var_6474_transpose_y_0, x = var_6473_cast_fp16, y = v_33_cast_fp16)[name = string("op_6474_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_65_cast_fp16 = add(x = var_6468_cast_fp16, y = var_6474_cast_fp16)[name = string("mixed_65_cast_fp16")];
            tensor<int32, [4]> var_6491_begin_0 = const()[name = string("op_6491_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6491_end_0 = const()[name = string("op_6491_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6491_end_mask_0 = const()[name = string("op_6491_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6491_cast_fp16 = slice_by_index(begin = var_6491_begin_0, end = var_6491_end_0, end_mask = var_6491_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6491_cast_fp16")];
            tensor<int32, [4]> var_6506_begin_0 = const()[name = string("op_6506_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6506_end_0 = const()[name = string("op_6506_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6506_end_mask_0 = const()[name = string("op_6506_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6506_cast_fp16 = slice_by_index(begin = var_6506_begin_0, end = var_6506_end_0, end_mask = var_6506_end_mask_0, x = mixed_65_cast_fp16)[name = string("op_6506_cast_fp16")];
            int32 var_6508 = const()[name = string("op_6508"), val = int32(1)];
            bool mixed_67_interleave_0 = const()[name = string("mixed_67_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_67_cast_fp16 = concat(axis = var_6508, interleave = mixed_67_interleave_0, values = (var_6491_cast_fp16, var_6506_cast_fp16))[name = string("mixed_67_cast_fp16")];
            tensor<int32, [4]> var_6512_perm_0 = const()[name = string("op_6512_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6516 = const()[name = string("op_6516"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6512_cast_fp16 = transpose(perm = var_6512_perm_0, x = mixed_67_cast_fp16)[name = string("transpose_55")];
            tensor<fp16, [1, 16, 2048]> input_163_cast_fp16 = reshape(shape = var_6516, x = var_6512_cast_fp16)[name = string("input_163_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_16_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(567597888))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569695104))))[name = string("model_layers_16_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_115_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_self_attn_o_proj_weight_to_fp16_quantized, x = input_163_cast_fp16)[name = string("linear_115_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_203_cast_fp16 = add(x = x_193_cast_fp16, y = linear_115_cast_fp16)[name = string("x_203_cast_fp16")];
            fp16 var_6527_promoted_to_fp16 = const()[name = string("op_6527_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6529_cast_fp16 = pow(x = x_203_cast_fp16, y = var_6527_promoted_to_fp16)[name = string("op_6529_cast_fp16")];
            tensor<int32, [1]> var_6531_axes_0 = const()[name = string("op_6531_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6531_keep_dims_0 = const()[name = string("op_6531_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6531_cast_fp16 = reduce_mean(axes = var_6531_axes_0, keep_dims = var_6531_keep_dims_0, x = var_6529_cast_fp16)[name = string("op_6531_cast_fp16")];
            fp16 var_6532_to_fp16 = const()[name = string("op_6532_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6533_cast_fp16 = add(x = var_6531_cast_fp16, y = var_6532_to_fp16)[name = string("op_6533_cast_fp16")];
            fp32 var_6534_epsilon_0 = const()[name = string("op_6534_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6534_cast_fp16 = rsqrt(epsilon = var_6534_epsilon_0, x = var_6533_cast_fp16)[name = string("op_6534_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6535_cast_fp16 = mul(x = x_203_cast_fp16, y = var_6534_cast_fp16)[name = string("op_6535_cast_fp16")];
            tensor<fp16, [1024]> model_layers_16_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_16_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569697216)))];
            tensor<fp16, [1, 16, 1024]> input_165_cast_fp16 = mul(x = var_6535_cast_fp16, y = model_layers_16_post_attention_layernorm_weight_to_fp16)[name = string("input_165_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(569699328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572845120))))[name = string("model_layers_16_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_116_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_gate_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_116_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6540_cast_fp16 = silu(x = linear_116_cast_fp16)[name = string("op_6540_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_16_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(572851328))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(575997120))))[name = string("model_layers_16_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_117_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_16_mlp_up_proj_weight_to_fp16_quantized, x = input_165_cast_fp16)[name = string("linear_117_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_169_cast_fp16 = mul(x = var_6540_cast_fp16, y = linear_117_cast_fp16)[name = string("input_169_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_16_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(576003328))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579149120))))[name = string("model_layers_16_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_118_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_16_mlp_down_proj_weight_to_fp16_quantized, x = input_169_cast_fp16)[name = string("linear_118_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_205_cast_fp16 = add(x = x_203_cast_fp16, y = linear_118_cast_fp16)[name = string("x_205_cast_fp16")];
            fp16 var_6554_promoted_to_fp16 = const()[name = string("op_6554_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6556_cast_fp16 = pow(x = x_205_cast_fp16, y = var_6554_promoted_to_fp16)[name = string("op_6556_cast_fp16")];
            tensor<int32, [1]> var_6558_axes_0 = const()[name = string("op_6558_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6558_keep_dims_0 = const()[name = string("op_6558_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6558_cast_fp16 = reduce_mean(axes = var_6558_axes_0, keep_dims = var_6558_keep_dims_0, x = var_6556_cast_fp16)[name = string("op_6558_cast_fp16")];
            fp16 var_6559_to_fp16 = const()[name = string("op_6559_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6560_cast_fp16 = add(x = var_6558_cast_fp16, y = var_6559_to_fp16)[name = string("op_6560_cast_fp16")];
            fp32 var_6561_epsilon_0 = const()[name = string("op_6561_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6561_cast_fp16 = rsqrt(epsilon = var_6561_epsilon_0, x = var_6560_cast_fp16)[name = string("op_6561_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6562_cast_fp16 = mul(x = x_205_cast_fp16, y = var_6561_cast_fp16)[name = string("op_6562_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579151232)))];
            tensor<fp16, [1, 16, 1024]> input_171_cast_fp16 = mul(x = var_6562_cast_fp16, y = model_layers_17_input_layernorm_weight_to_fp16)[name = string("input_171_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_17_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(579153344))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581250560))))[name = string("model_layers_17_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_119_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_17_self_attn_q_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_119_cast_fp16")];
            tensor<int32, [4]> var_6571 = const()[name = string("op_6571"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6572_cast_fp16 = reshape(shape = var_6571, x = linear_119_cast_fp16)[name = string("op_6572_cast_fp16")];
            tensor<int32, [4]> x_207_perm_0 = const()[name = string("x_207_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6580_promoted_to_fp16 = const()[name = string("op_6580_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_207_cast_fp16 = transpose(perm = x_207_perm_0, x = var_6572_cast_fp16)[name = string("transpose_54")];
            tensor<fp16, [1, 16, 16, 128]> var_6582_cast_fp16 = pow(x = x_207_cast_fp16, y = var_6580_promoted_to_fp16)[name = string("op_6582_cast_fp16")];
            tensor<int32, [1]> var_6584_axes_0 = const()[name = string("op_6584_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6584_keep_dims_0 = const()[name = string("op_6584_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6584_cast_fp16 = reduce_mean(axes = var_6584_axes_0, keep_dims = var_6584_keep_dims_0, x = var_6582_cast_fp16)[name = string("op_6584_cast_fp16")];
            fp16 var_6585_to_fp16 = const()[name = string("op_6585_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6586_cast_fp16 = add(x = var_6584_cast_fp16, y = var_6585_to_fp16)[name = string("op_6586_cast_fp16")];
            fp32 var_6587_epsilon_0 = const()[name = string("op_6587_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6587_cast_fp16 = rsqrt(epsilon = var_6587_epsilon_0, x = var_6586_cast_fp16)[name = string("op_6587_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6588_cast_fp16 = mul(x = x_207_cast_fp16, y = var_6587_cast_fp16)[name = string("op_6588_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581254720)))];
            tensor<fp16, [1, 16, 16, 128]> x_211_cast_fp16 = mul(x = var_6588_cast_fp16, y = model_layers_17_self_attn_q_norm_weight_to_fp16)[name = string("x_211_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(581255040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582303680))))[name = string("model_layers_17_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_120_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_k_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_120_cast_fp16")];
            tensor<int32, [4]> var_6597 = const()[name = string("op_6597"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6598_cast_fp16 = reshape(shape = var_6597, x = linear_120_cast_fp16)[name = string("op_6598_cast_fp16")];
            tensor<int32, [4]> x_209_perm_0 = const()[name = string("x_209_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6606_promoted_to_fp16 = const()[name = string("op_6606_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_209_cast_fp16 = transpose(perm = x_209_perm_0, x = var_6598_cast_fp16)[name = string("transpose_53")];
            tensor<fp16, [1, 8, 16, 128]> var_6608_cast_fp16 = pow(x = x_209_cast_fp16, y = var_6606_promoted_to_fp16)[name = string("op_6608_cast_fp16")];
            tensor<int32, [1]> var_6610_axes_0 = const()[name = string("op_6610_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6610_keep_dims_0 = const()[name = string("op_6610_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6610_cast_fp16 = reduce_mean(axes = var_6610_axes_0, keep_dims = var_6610_keep_dims_0, x = var_6608_cast_fp16)[name = string("op_6610_cast_fp16")];
            fp16 var_6611_to_fp16 = const()[name = string("op_6611_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6612_cast_fp16 = add(x = var_6610_cast_fp16, y = var_6611_to_fp16)[name = string("op_6612_cast_fp16")];
            fp32 var_6613_epsilon_0 = const()[name = string("op_6613_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6613_cast_fp16 = rsqrt(epsilon = var_6613_epsilon_0, x = var_6612_cast_fp16)[name = string("op_6613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6614_cast_fp16 = mul(x = x_209_cast_fp16, y = var_6613_cast_fp16)[name = string("op_6614_cast_fp16")];
            tensor<fp16, [128]> model_layers_17_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_17_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582305792)))];
            tensor<fp16, [1, 8, 16, 128]> x_213_cast_fp16 = mul(x = var_6614_cast_fp16, y = model_layers_17_self_attn_k_norm_weight_to_fp16)[name = string("x_213_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_17_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(582306112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583354752))))[name = string("model_layers_17_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_121_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_v_proj_weight_to_fp16_quantized, x = input_171_cast_fp16)[name = string("linear_121_cast_fp16")];
            tensor<int32, [4]> var_6623 = const()[name = string("op_6623"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6624_cast_fp16 = reshape(shape = var_6623, x = linear_121_cast_fp16)[name = string("op_6624_cast_fp16")];
            tensor<int32, [4]> v_35_perm_0 = const()[name = string("v_35_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_69_begin_0 = const()[name = string("first_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_69_end_0 = const()[name = string("first_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_69_end_mask_0 = const()[name = string("first_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_69_cast_fp16 = slice_by_index(begin = first_69_begin_0, end = first_69_end_0, end_mask = first_69_end_mask_0, x = x_211_cast_fp16)[name = string("first_69_cast_fp16")];
            tensor<int32, [4]> second_69_begin_0 = const()[name = string("second_69_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_69_end_0 = const()[name = string("second_69_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_69_end_mask_0 = const()[name = string("second_69_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_69_cast_fp16 = slice_by_index(begin = second_69_begin_0, end = second_69_end_0, end_mask = second_69_end_mask_0, x = x_211_cast_fp16)[name = string("second_69_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6638_cast_fp16 = mul(x = first_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6638_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6639_cast_fp16 = mul(x = second_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6639_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6641_cast_fp16 = sub(x = var_6638_cast_fp16, y = var_6639_cast_fp16)[name = string("op_6641_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6642_cast_fp16 = mul(x = second_69_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6642_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6643_cast_fp16 = mul(x = first_69_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6643_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6645_cast_fp16 = add(x = var_6642_cast_fp16, y = var_6643_cast_fp16)[name = string("op_6645_cast_fp16")];
            int32 var_6647 = const()[name = string("op_6647"), val = int32(-1)];
            bool q_35_interleave_0 = const()[name = string("q_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_35_cast_fp16 = concat(axis = var_6647, interleave = q_35_interleave_0, values = (var_6641_cast_fp16, var_6645_cast_fp16))[name = string("q_35_cast_fp16")];
            tensor<int32, [4]> first_71_begin_0 = const()[name = string("first_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_71_end_0 = const()[name = string("first_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_71_end_mask_0 = const()[name = string("first_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_71_cast_fp16 = slice_by_index(begin = first_71_begin_0, end = first_71_end_0, end_mask = first_71_end_mask_0, x = x_213_cast_fp16)[name = string("first_71_cast_fp16")];
            tensor<int32, [4]> second_71_begin_0 = const()[name = string("second_71_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_71_end_0 = const()[name = string("second_71_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_71_end_mask_0 = const()[name = string("second_71_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_71_cast_fp16 = slice_by_index(begin = second_71_begin_0, end = second_71_end_0, end_mask = second_71_end_mask_0, x = x_213_cast_fp16)[name = string("second_71_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6659_cast_fp16 = mul(x = first_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6659_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6660_cast_fp16 = mul(x = second_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6660_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6662_cast_fp16 = sub(x = var_6659_cast_fp16, y = var_6660_cast_fp16)[name = string("op_6662_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6663_cast_fp16 = mul(x = second_71_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6663_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6664_cast_fp16 = mul(x = first_71_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6664_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6666_cast_fp16 = add(x = var_6663_cast_fp16, y = var_6664_cast_fp16)[name = string("op_6666_cast_fp16")];
            int32 var_6668 = const()[name = string("op_6668"), val = int32(-1)];
            bool k_35_interleave_0 = const()[name = string("k_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_35_cast_fp16 = concat(axis = var_6668, interleave = k_35_interleave_0, values = (var_6662_cast_fp16, var_6666_cast_fp16))[name = string("k_35_cast_fp16")];
            tensor<int32, [4]> var_6679_begin_0 = const()[name = string("op_6679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6679_end_0 = const()[name = string("op_6679_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6679_end_mask_0 = const()[name = string("op_6679_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6679_stride_0 = const()[name = string("op_6679_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6679_cast_fp16 = slice_by_index(begin = var_6679_begin_0, end = var_6679_end_0, end_mask = var_6679_end_mask_0, stride = var_6679_stride_0, x = q_35_cast_fp16)[name = string("op_6679_cast_fp16")];
            tensor<int32, [4]> var_6689_begin_0 = const()[name = string("op_6689_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6689_end_0 = const()[name = string("op_6689_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6689_end_mask_0 = const()[name = string("op_6689_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6689_stride_0 = const()[name = string("op_6689_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6689_cast_fp16 = slice_by_index(begin = var_6689_begin_0, end = var_6689_end_0, end_mask = var_6689_end_mask_0, stride = var_6689_stride_0, x = q_35_cast_fp16)[name = string("op_6689_cast_fp16")];
            int32 var_6691 = const()[name = string("op_6691"), val = int32(2)];
            bool stacked_35_interleave_0 = const()[name = string("stacked_35_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_35_cast_fp16 = concat(axis = var_6691, interleave = stacked_35_interleave_0, values = (var_6679_cast_fp16, var_6689_cast_fp16))[name = string("stacked_35_cast_fp16")];
            tensor<int32, [4]> var_6697_begin_0 = const()[name = string("op_6697_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6697_end_0 = const()[name = string("op_6697_end_0"), val = tensor<int32, [4]>([18, 8, 528, 128])];
            tensor<bool, [4]> var_6697_end_mask_0 = const()[name = string("op_6697_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6697_cast_fp16 = slice_by_index(begin = var_6697_begin_0, end = var_6697_end_0, end_mask = var_6697_end_mask_0, x = read_state_0)[name = string("op_6697_cast_fp16")];
            tensor<int32, [4]> var_6707_begin_0 = const()[name = string("op_6707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6707_end_0 = const()[name = string("op_6707_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6707_end_mask_0 = const()[name = string("op_6707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6707_cast_fp16 = slice_by_index(begin = var_6707_begin_0, end = var_6707_end_0, end_mask = var_6707_end_mask_0, x = var_6697_cast_fp16)[name = string("op_6707_cast_fp16")];
            bool var_6716_transpose_x_1 = const()[name = string("op_6716_transpose_x_1"), val = bool(false)];
            bool var_6716_transpose_y_1 = const()[name = string("op_6716_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_6716_cast_fp16 = matmul(transpose_x = var_6716_transpose_x_1, transpose_y = var_6716_transpose_y_1, x = var_6707_cast_fp16, y = stacked_35_cast_fp16)[name = string("op_6716_cast_fp16")];
            tensor<int32, [4]> var_6719_perm_0 = const()[name = string("op_6719_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_6720_to_fp16 = const()[name = string("op_6720_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_6719_cast_fp16 = transpose(perm = var_6719_perm_0, x = var_6716_cast_fp16)[name = string("transpose_52")];
            tensor<fp16, [1, 8, 32, 512]> var_6721_cast_fp16 = mul(x = var_6719_cast_fp16, y = var_6720_to_fp16)[name = string("op_6721_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_35_cast_fp16 = add(x = var_6721_cast_fp16, y = cacheMask)[name = string("cache_weights_35_cast_fp16")];
            bool var_6727_transpose_x_1 = const()[name = string("op_6727_transpose_x_1"), val = bool(false)];
            bool var_6727_transpose_y_1 = const()[name = string("op_6727_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_6727_cast_fp16 = matmul(transpose_x = var_6727_transpose_x_1, transpose_y = var_6727_transpose_y_1, x = stacked_35_cast_fp16, y = k_35_cast_fp16)[name = string("op_6727_cast_fp16")];
            fp16 var_6728_to_fp16 = const()[name = string("op_6728_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_6729_cast_fp16 = mul(x = var_6727_cast_fp16, y = var_6728_to_fp16)[name = string("op_6729_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_35_cast_fp16 = add(x = var_6729_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_35_cast_fp16")];
            int32 var_6733 = const()[name = string("op_6733"), val = int32(-1)];
            bool var_6734_interleave_0 = const()[name = string("op_6734_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_6734_cast_fp16 = concat(axis = var_6733, interleave = var_6734_interleave_0, values = (cache_weights_35_cast_fp16, call_weights_35_cast_fp16))[name = string("op_6734_cast_fp16")];
            int32 var_6735 = const()[name = string("op_6735"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_35_cast_fp16 = softmax(axis = var_6735, x = var_6734_cast_fp16)[name = string("shares_35_cast_fp16")];
            tensor<int32, [4]> var_6742_begin_0 = const()[name = string("op_6742_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6742_end_0 = const()[name = string("op_6742_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_6742_end_mask_0 = const()[name = string("op_6742_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_6742_cast_fp16 = slice_by_index(begin = var_6742_begin_0, end = var_6742_end_0, end_mask = var_6742_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6742_cast_fp16")];
            tensor<int32, [4]> var_6747_begin_0 = const()[name = string("op_6747_begin_0"), val = tensor<int32, [4]>([17, 0, 0, 0])];
            tensor<int32, [4]> var_6747_end_0 = const()[name = string("op_6747_end_0"), val = tensor<int32, [4]>([18, 8, 528, 128])];
            tensor<bool, [4]> var_6747_end_mask_0 = const()[name = string("op_6747_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6747_cast_fp16 = slice_by_index(begin = var_6747_begin_0, end = var_6747_end_0, end_mask = var_6747_end_mask_0, x = read_state_1)[name = string("op_6747_cast_fp16")];
            tensor<int32, [4]> var_6757_begin_0 = const()[name = string("op_6757_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6757_end_0 = const()[name = string("op_6757_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_6757_end_mask_0 = const()[name = string("op_6757_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_6757_cast_fp16 = slice_by_index(begin = var_6757_begin_0, end = var_6757_end_0, end_mask = var_6757_end_mask_0, x = var_6747_cast_fp16)[name = string("op_6757_cast_fp16")];
            bool var_6763_transpose_x_0 = const()[name = string("op_6763_transpose_x_0"), val = bool(false)];
            bool var_6763_transpose_y_0 = const()[name = string("op_6763_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_6763_cast_fp16 = matmul(transpose_x = var_6763_transpose_x_0, transpose_y = var_6763_transpose_y_0, x = var_6742_cast_fp16, y = var_6757_cast_fp16)[name = string("op_6763_cast_fp16")];
            tensor<int32, [4]> var_6768_begin_0 = const()[name = string("op_6768_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_6768_end_0 = const()[name = string("op_6768_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_6768_end_mask_0 = const()[name = string("op_6768_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_6768_cast_fp16 = slice_by_index(begin = var_6768_begin_0, end = var_6768_end_0, end_mask = var_6768_end_mask_0, x = shares_35_cast_fp16)[name = string("op_6768_cast_fp16")];
            bool var_6769_transpose_x_0 = const()[name = string("op_6769_transpose_x_0"), val = bool(false)];
            bool var_6769_transpose_y_0 = const()[name = string("op_6769_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_35_cast_fp16 = transpose(perm = v_35_perm_0, x = var_6624_cast_fp16)[name = string("transpose_51")];
            tensor<fp16, [1, 8, 32, 128]> var_6769_cast_fp16 = matmul(transpose_x = var_6769_transpose_x_0, transpose_y = var_6769_transpose_y_0, x = var_6768_cast_fp16, y = v_35_cast_fp16)[name = string("op_6769_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_69_cast_fp16 = add(x = var_6763_cast_fp16, y = var_6769_cast_fp16)[name = string("mixed_69_cast_fp16")];
            tensor<int32, [4]> var_6786_begin_0 = const()[name = string("op_6786_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6786_end_0 = const()[name = string("op_6786_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_6786_end_mask_0 = const()[name = string("op_6786_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6786_cast_fp16 = slice_by_index(begin = var_6786_begin_0, end = var_6786_end_0, end_mask = var_6786_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6786_cast_fp16")];
            tensor<int32, [4]> var_6801_begin_0 = const()[name = string("op_6801_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_6801_end_0 = const()[name = string("op_6801_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_6801_end_mask_0 = const()[name = string("op_6801_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_6801_cast_fp16 = slice_by_index(begin = var_6801_begin_0, end = var_6801_end_0, end_mask = var_6801_end_mask_0, x = mixed_69_cast_fp16)[name = string("op_6801_cast_fp16")];
            int32 var_6803 = const()[name = string("op_6803"), val = int32(1)];
            bool mixed_71_interleave_0 = const()[name = string("mixed_71_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_71_cast_fp16 = concat(axis = var_6803, interleave = mixed_71_interleave_0, values = (var_6786_cast_fp16, var_6801_cast_fp16))[name = string("mixed_71_cast_fp16")];
            tensor<int32, [4]> var_6807_perm_0 = const()[name = string("op_6807_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_6811 = const()[name = string("op_6811"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_6807_cast_fp16 = transpose(perm = var_6807_perm_0, x = mixed_71_cast_fp16)[name = string("transpose_50")];
            tensor<fp16, [1, 16, 2048]> input_173_cast_fp16 = reshape(shape = var_6811, x = var_6807_cast_fp16)[name = string("input_173_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_17_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(583356864))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585454080))))[name = string("model_layers_17_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_122_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_self_attn_o_proj_weight_to_fp16_quantized, x = input_173_cast_fp16)[name = string("linear_122_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_215_cast_fp16 = add(x = x_205_cast_fp16, y = linear_122_cast_fp16)[name = string("x_215_cast_fp16")];
            fp16 var_6822_promoted_to_fp16 = const()[name = string("op_6822_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6824_cast_fp16 = pow(x = x_215_cast_fp16, y = var_6822_promoted_to_fp16)[name = string("op_6824_cast_fp16")];
            tensor<int32, [1]> var_6826_axes_0 = const()[name = string("op_6826_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6826_keep_dims_0 = const()[name = string("op_6826_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6826_cast_fp16 = reduce_mean(axes = var_6826_axes_0, keep_dims = var_6826_keep_dims_0, x = var_6824_cast_fp16)[name = string("op_6826_cast_fp16")];
            fp16 var_6827_to_fp16 = const()[name = string("op_6827_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6828_cast_fp16 = add(x = var_6826_cast_fp16, y = var_6827_to_fp16)[name = string("op_6828_cast_fp16")];
            fp32 var_6829_epsilon_0 = const()[name = string("op_6829_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6829_cast_fp16 = rsqrt(epsilon = var_6829_epsilon_0, x = var_6828_cast_fp16)[name = string("op_6829_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6830_cast_fp16 = mul(x = x_215_cast_fp16, y = var_6829_cast_fp16)[name = string("op_6830_cast_fp16")];
            tensor<fp16, [1024]> model_layers_17_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_17_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585456192)))];
            tensor<fp16, [1, 16, 1024]> input_175_cast_fp16 = mul(x = var_6830_cast_fp16, y = model_layers_17_post_attention_layernorm_weight_to_fp16)[name = string("input_175_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(585458304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588604096))))[name = string("model_layers_17_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_123_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_gate_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_123_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_6835_cast_fp16 = silu(x = linear_123_cast_fp16)[name = string("op_6835_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_17_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588610304))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591756096))))[name = string("model_layers_17_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_124_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_17_mlp_up_proj_weight_to_fp16_quantized, x = input_175_cast_fp16)[name = string("linear_124_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_179_cast_fp16 = mul(x = var_6835_cast_fp16, y = linear_124_cast_fp16)[name = string("input_179_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_17_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(591762304))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594908096))))[name = string("model_layers_17_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_125_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_17_mlp_down_proj_weight_to_fp16_quantized, x = input_179_cast_fp16)[name = string("linear_125_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_217_cast_fp16 = add(x = x_215_cast_fp16, y = linear_125_cast_fp16)[name = string("x_217_cast_fp16")];
            fp16 var_6849_promoted_to_fp16 = const()[name = string("op_6849_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_6851_cast_fp16 = pow(x = x_217_cast_fp16, y = var_6849_promoted_to_fp16)[name = string("op_6851_cast_fp16")];
            tensor<int32, [1]> var_6853_axes_0 = const()[name = string("op_6853_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6853_keep_dims_0 = const()[name = string("op_6853_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_6853_cast_fp16 = reduce_mean(axes = var_6853_axes_0, keep_dims = var_6853_keep_dims_0, x = var_6851_cast_fp16)[name = string("op_6853_cast_fp16")];
            fp16 var_6854_to_fp16 = const()[name = string("op_6854_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_6855_cast_fp16 = add(x = var_6853_cast_fp16, y = var_6854_to_fp16)[name = string("op_6855_cast_fp16")];
            fp32 var_6856_epsilon_0 = const()[name = string("op_6856_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_6856_cast_fp16 = rsqrt(epsilon = var_6856_epsilon_0, x = var_6855_cast_fp16)[name = string("op_6856_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_6857_cast_fp16 = mul(x = x_217_cast_fp16, y = var_6856_cast_fp16)[name = string("op_6857_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594910208)))];
            tensor<fp16, [1, 16, 1024]> input_181_cast_fp16 = mul(x = var_6857_cast_fp16, y = model_layers_18_input_layernorm_weight_to_fp16)[name = string("input_181_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_18_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(594912320))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597009536))))[name = string("model_layers_18_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_126_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_18_self_attn_q_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_126_cast_fp16")];
            tensor<int32, [4]> var_6866 = const()[name = string("op_6866"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_6867_cast_fp16 = reshape(shape = var_6866, x = linear_126_cast_fp16)[name = string("op_6867_cast_fp16")];
            tensor<int32, [4]> x_219_perm_0 = const()[name = string("x_219_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6875_promoted_to_fp16 = const()[name = string("op_6875_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_219_cast_fp16 = transpose(perm = x_219_perm_0, x = var_6867_cast_fp16)[name = string("transpose_49")];
            tensor<fp16, [1, 16, 16, 128]> var_6877_cast_fp16 = pow(x = x_219_cast_fp16, y = var_6875_promoted_to_fp16)[name = string("op_6877_cast_fp16")];
            tensor<int32, [1]> var_6879_axes_0 = const()[name = string("op_6879_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6879_keep_dims_0 = const()[name = string("op_6879_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_6879_cast_fp16 = reduce_mean(axes = var_6879_axes_0, keep_dims = var_6879_keep_dims_0, x = var_6877_cast_fp16)[name = string("op_6879_cast_fp16")];
            fp16 var_6880_to_fp16 = const()[name = string("op_6880_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_6881_cast_fp16 = add(x = var_6879_cast_fp16, y = var_6880_to_fp16)[name = string("op_6881_cast_fp16")];
            fp32 var_6882_epsilon_0 = const()[name = string("op_6882_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_6882_cast_fp16 = rsqrt(epsilon = var_6882_epsilon_0, x = var_6881_cast_fp16)[name = string("op_6882_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_6883_cast_fp16 = mul(x = x_219_cast_fp16, y = var_6882_cast_fp16)[name = string("op_6883_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597013696)))];
            tensor<fp16, [1, 16, 16, 128]> x_223_cast_fp16 = mul(x = var_6883_cast_fp16, y = model_layers_18_self_attn_q_norm_weight_to_fp16)[name = string("x_223_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(597014016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598062656))))[name = string("model_layers_18_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_127_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_k_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_127_cast_fp16")];
            tensor<int32, [4]> var_6892 = const()[name = string("op_6892"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6893_cast_fp16 = reshape(shape = var_6892, x = linear_127_cast_fp16)[name = string("op_6893_cast_fp16")];
            tensor<int32, [4]> x_221_perm_0 = const()[name = string("x_221_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_6901_promoted_to_fp16 = const()[name = string("op_6901_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_221_cast_fp16 = transpose(perm = x_221_perm_0, x = var_6893_cast_fp16)[name = string("transpose_48")];
            tensor<fp16, [1, 8, 16, 128]> var_6903_cast_fp16 = pow(x = x_221_cast_fp16, y = var_6901_promoted_to_fp16)[name = string("op_6903_cast_fp16")];
            tensor<int32, [1]> var_6905_axes_0 = const()[name = string("op_6905_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_6905_keep_dims_0 = const()[name = string("op_6905_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_6905_cast_fp16 = reduce_mean(axes = var_6905_axes_0, keep_dims = var_6905_keep_dims_0, x = var_6903_cast_fp16)[name = string("op_6905_cast_fp16")];
            fp16 var_6906_to_fp16 = const()[name = string("op_6906_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_6907_cast_fp16 = add(x = var_6905_cast_fp16, y = var_6906_to_fp16)[name = string("op_6907_cast_fp16")];
            fp32 var_6908_epsilon_0 = const()[name = string("op_6908_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_6908_cast_fp16 = rsqrt(epsilon = var_6908_epsilon_0, x = var_6907_cast_fp16)[name = string("op_6908_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_6909_cast_fp16 = mul(x = x_221_cast_fp16, y = var_6908_cast_fp16)[name = string("op_6909_cast_fp16")];
            tensor<fp16, [128]> model_layers_18_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_18_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598064768)))];
            tensor<fp16, [1, 8, 16, 128]> x_225_cast_fp16 = mul(x = var_6909_cast_fp16, y = model_layers_18_self_attn_k_norm_weight_to_fp16)[name = string("x_225_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_18_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(598065088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599113728))))[name = string("model_layers_18_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_128_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_v_proj_weight_to_fp16_quantized, x = input_181_cast_fp16)[name = string("linear_128_cast_fp16")];
            tensor<int32, [4]> var_6918 = const()[name = string("op_6918"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_6919_cast_fp16 = reshape(shape = var_6918, x = linear_128_cast_fp16)[name = string("op_6919_cast_fp16")];
            tensor<int32, [4]> v_37_perm_0 = const()[name = string("v_37_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_73_begin_0 = const()[name = string("first_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_73_end_0 = const()[name = string("first_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_73_end_mask_0 = const()[name = string("first_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_73_cast_fp16 = slice_by_index(begin = first_73_begin_0, end = first_73_end_0, end_mask = first_73_end_mask_0, x = x_223_cast_fp16)[name = string("first_73_cast_fp16")];
            tensor<int32, [4]> second_73_begin_0 = const()[name = string("second_73_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_73_end_0 = const()[name = string("second_73_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_73_end_mask_0 = const()[name = string("second_73_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_73_cast_fp16 = slice_by_index(begin = second_73_begin_0, end = second_73_end_0, end_mask = second_73_end_mask_0, x = x_223_cast_fp16)[name = string("second_73_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6933_cast_fp16 = mul(x = first_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6933_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6934_cast_fp16 = mul(x = second_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6934_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6936_cast_fp16 = sub(x = var_6933_cast_fp16, y = var_6934_cast_fp16)[name = string("op_6936_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6937_cast_fp16 = mul(x = second_73_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6937_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6938_cast_fp16 = mul(x = first_73_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6938_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_6940_cast_fp16 = add(x = var_6937_cast_fp16, y = var_6938_cast_fp16)[name = string("op_6940_cast_fp16")];
            int32 var_6942 = const()[name = string("op_6942"), val = int32(-1)];
            bool q_37_interleave_0 = const()[name = string("q_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_37_cast_fp16 = concat(axis = var_6942, interleave = q_37_interleave_0, values = (var_6936_cast_fp16, var_6940_cast_fp16))[name = string("q_37_cast_fp16")];
            tensor<int32, [4]> first_75_begin_0 = const()[name = string("first_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_75_end_0 = const()[name = string("first_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_75_end_mask_0 = const()[name = string("first_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_75_cast_fp16 = slice_by_index(begin = first_75_begin_0, end = first_75_end_0, end_mask = first_75_end_mask_0, x = x_225_cast_fp16)[name = string("first_75_cast_fp16")];
            tensor<int32, [4]> second_75_begin_0 = const()[name = string("second_75_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_75_end_0 = const()[name = string("second_75_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_75_end_mask_0 = const()[name = string("second_75_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_75_cast_fp16 = slice_by_index(begin = second_75_begin_0, end = second_75_end_0, end_mask = second_75_end_mask_0, x = x_225_cast_fp16)[name = string("second_75_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6954_cast_fp16 = mul(x = first_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6954_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6955_cast_fp16 = mul(x = second_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6955_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6957_cast_fp16 = sub(x = var_6954_cast_fp16, y = var_6955_cast_fp16)[name = string("op_6957_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6958_cast_fp16 = mul(x = second_75_cast_fp16, y = cos_1_cast_fp16)[name = string("op_6958_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6959_cast_fp16 = mul(x = first_75_cast_fp16, y = sin_1_cast_fp16)[name = string("op_6959_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_6961_cast_fp16 = add(x = var_6958_cast_fp16, y = var_6959_cast_fp16)[name = string("op_6961_cast_fp16")];
            int32 var_6963 = const()[name = string("op_6963"), val = int32(-1)];
            bool k_37_interleave_0 = const()[name = string("k_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_37_cast_fp16 = concat(axis = var_6963, interleave = k_37_interleave_0, values = (var_6957_cast_fp16, var_6961_cast_fp16))[name = string("k_37_cast_fp16")];
            tensor<int32, [4]> var_6974_begin_0 = const()[name = string("op_6974_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_6974_end_0 = const()[name = string("op_6974_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6974_end_mask_0 = const()[name = string("op_6974_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6974_stride_0 = const()[name = string("op_6974_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6974_cast_fp16 = slice_by_index(begin = var_6974_begin_0, end = var_6974_end_0, end_mask = var_6974_end_mask_0, stride = var_6974_stride_0, x = q_37_cast_fp16)[name = string("op_6974_cast_fp16")];
            tensor<int32, [4]> var_6984_begin_0 = const()[name = string("op_6984_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_6984_end_0 = const()[name = string("op_6984_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_6984_end_mask_0 = const()[name = string("op_6984_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_6984_stride_0 = const()[name = string("op_6984_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_6984_cast_fp16 = slice_by_index(begin = var_6984_begin_0, end = var_6984_end_0, end_mask = var_6984_end_mask_0, stride = var_6984_stride_0, x = q_37_cast_fp16)[name = string("op_6984_cast_fp16")];
            int32 var_6986 = const()[name = string("op_6986"), val = int32(2)];
            bool stacked_37_interleave_0 = const()[name = string("stacked_37_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_37_cast_fp16 = concat(axis = var_6986, interleave = stacked_37_interleave_0, values = (var_6974_cast_fp16, var_6984_cast_fp16))[name = string("stacked_37_cast_fp16")];
            tensor<int32, [4]> var_6992_begin_0 = const()[name = string("op_6992_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_6992_end_0 = const()[name = string("op_6992_end_0"), val = tensor<int32, [4]>([19, 8, 528, 128])];
            tensor<bool, [4]> var_6992_end_mask_0 = const()[name = string("op_6992_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_6992_cast_fp16 = slice_by_index(begin = var_6992_begin_0, end = var_6992_end_0, end_mask = var_6992_end_mask_0, x = read_state_0)[name = string("op_6992_cast_fp16")];
            tensor<int32, [4]> var_7002_begin_0 = const()[name = string("op_7002_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7002_end_0 = const()[name = string("op_7002_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7002_end_mask_0 = const()[name = string("op_7002_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7002_cast_fp16 = slice_by_index(begin = var_7002_begin_0, end = var_7002_end_0, end_mask = var_7002_end_mask_0, x = var_6992_cast_fp16)[name = string("op_7002_cast_fp16")];
            bool var_7011_transpose_x_1 = const()[name = string("op_7011_transpose_x_1"), val = bool(false)];
            bool var_7011_transpose_y_1 = const()[name = string("op_7011_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7011_cast_fp16 = matmul(transpose_x = var_7011_transpose_x_1, transpose_y = var_7011_transpose_y_1, x = var_7002_cast_fp16, y = stacked_37_cast_fp16)[name = string("op_7011_cast_fp16")];
            tensor<int32, [4]> var_7014_perm_0 = const()[name = string("op_7014_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7015_to_fp16 = const()[name = string("op_7015_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7014_cast_fp16 = transpose(perm = var_7014_perm_0, x = var_7011_cast_fp16)[name = string("transpose_47")];
            tensor<fp16, [1, 8, 32, 512]> var_7016_cast_fp16 = mul(x = var_7014_cast_fp16, y = var_7015_to_fp16)[name = string("op_7016_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_37_cast_fp16 = add(x = var_7016_cast_fp16, y = cacheMask)[name = string("cache_weights_37_cast_fp16")];
            bool var_7022_transpose_x_1 = const()[name = string("op_7022_transpose_x_1"), val = bool(false)];
            bool var_7022_transpose_y_1 = const()[name = string("op_7022_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7022_cast_fp16 = matmul(transpose_x = var_7022_transpose_x_1, transpose_y = var_7022_transpose_y_1, x = stacked_37_cast_fp16, y = k_37_cast_fp16)[name = string("op_7022_cast_fp16")];
            fp16 var_7023_to_fp16 = const()[name = string("op_7023_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7024_cast_fp16 = mul(x = var_7022_cast_fp16, y = var_7023_to_fp16)[name = string("op_7024_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_37_cast_fp16 = add(x = var_7024_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_37_cast_fp16")];
            int32 var_7028 = const()[name = string("op_7028"), val = int32(-1)];
            bool var_7029_interleave_0 = const()[name = string("op_7029_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7029_cast_fp16 = concat(axis = var_7028, interleave = var_7029_interleave_0, values = (cache_weights_37_cast_fp16, call_weights_37_cast_fp16))[name = string("op_7029_cast_fp16")];
            int32 var_7030 = const()[name = string("op_7030"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_37_cast_fp16 = softmax(axis = var_7030, x = var_7029_cast_fp16)[name = string("shares_37_cast_fp16")];
            tensor<int32, [4]> var_7037_begin_0 = const()[name = string("op_7037_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7037_end_0 = const()[name = string("op_7037_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7037_end_mask_0 = const()[name = string("op_7037_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7037_cast_fp16 = slice_by_index(begin = var_7037_begin_0, end = var_7037_end_0, end_mask = var_7037_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7037_cast_fp16")];
            tensor<int32, [4]> var_7042_begin_0 = const()[name = string("op_7042_begin_0"), val = tensor<int32, [4]>([18, 0, 0, 0])];
            tensor<int32, [4]> var_7042_end_0 = const()[name = string("op_7042_end_0"), val = tensor<int32, [4]>([19, 8, 528, 128])];
            tensor<bool, [4]> var_7042_end_mask_0 = const()[name = string("op_7042_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7042_cast_fp16 = slice_by_index(begin = var_7042_begin_0, end = var_7042_end_0, end_mask = var_7042_end_mask_0, x = read_state_1)[name = string("op_7042_cast_fp16")];
            tensor<int32, [4]> var_7052_begin_0 = const()[name = string("op_7052_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7052_end_0 = const()[name = string("op_7052_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7052_end_mask_0 = const()[name = string("op_7052_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7052_cast_fp16 = slice_by_index(begin = var_7052_begin_0, end = var_7052_end_0, end_mask = var_7052_end_mask_0, x = var_7042_cast_fp16)[name = string("op_7052_cast_fp16")];
            bool var_7058_transpose_x_0 = const()[name = string("op_7058_transpose_x_0"), val = bool(false)];
            bool var_7058_transpose_y_0 = const()[name = string("op_7058_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7058_cast_fp16 = matmul(transpose_x = var_7058_transpose_x_0, transpose_y = var_7058_transpose_y_0, x = var_7037_cast_fp16, y = var_7052_cast_fp16)[name = string("op_7058_cast_fp16")];
            tensor<int32, [4]> var_7063_begin_0 = const()[name = string("op_7063_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7063_end_0 = const()[name = string("op_7063_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7063_end_mask_0 = const()[name = string("op_7063_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7063_cast_fp16 = slice_by_index(begin = var_7063_begin_0, end = var_7063_end_0, end_mask = var_7063_end_mask_0, x = shares_37_cast_fp16)[name = string("op_7063_cast_fp16")];
            bool var_7064_transpose_x_0 = const()[name = string("op_7064_transpose_x_0"), val = bool(false)];
            bool var_7064_transpose_y_0 = const()[name = string("op_7064_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_37_cast_fp16 = transpose(perm = v_37_perm_0, x = var_6919_cast_fp16)[name = string("transpose_46")];
            tensor<fp16, [1, 8, 32, 128]> var_7064_cast_fp16 = matmul(transpose_x = var_7064_transpose_x_0, transpose_y = var_7064_transpose_y_0, x = var_7063_cast_fp16, y = v_37_cast_fp16)[name = string("op_7064_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_73_cast_fp16 = add(x = var_7058_cast_fp16, y = var_7064_cast_fp16)[name = string("mixed_73_cast_fp16")];
            tensor<int32, [4]> var_7081_begin_0 = const()[name = string("op_7081_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7081_end_0 = const()[name = string("op_7081_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7081_end_mask_0 = const()[name = string("op_7081_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7081_cast_fp16 = slice_by_index(begin = var_7081_begin_0, end = var_7081_end_0, end_mask = var_7081_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7081_cast_fp16")];
            tensor<int32, [4]> var_7096_begin_0 = const()[name = string("op_7096_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7096_end_0 = const()[name = string("op_7096_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7096_end_mask_0 = const()[name = string("op_7096_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7096_cast_fp16 = slice_by_index(begin = var_7096_begin_0, end = var_7096_end_0, end_mask = var_7096_end_mask_0, x = mixed_73_cast_fp16)[name = string("op_7096_cast_fp16")];
            int32 var_7098 = const()[name = string("op_7098"), val = int32(1)];
            bool mixed_75_interleave_0 = const()[name = string("mixed_75_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_75_cast_fp16 = concat(axis = var_7098, interleave = mixed_75_interleave_0, values = (var_7081_cast_fp16, var_7096_cast_fp16))[name = string("mixed_75_cast_fp16")];
            tensor<int32, [4]> var_7102_perm_0 = const()[name = string("op_7102_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7106 = const()[name = string("op_7106"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7102_cast_fp16 = transpose(perm = var_7102_perm_0, x = mixed_75_cast_fp16)[name = string("transpose_45")];
            tensor<fp16, [1, 16, 2048]> input_183_cast_fp16 = reshape(shape = var_7106, x = var_7102_cast_fp16)[name = string("input_183_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_18_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(599115840))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601213056))))[name = string("model_layers_18_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_129_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_self_attn_o_proj_weight_to_fp16_quantized, x = input_183_cast_fp16)[name = string("linear_129_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_227_cast_fp16 = add(x = x_217_cast_fp16, y = linear_129_cast_fp16)[name = string("x_227_cast_fp16")];
            fp16 var_7117_promoted_to_fp16 = const()[name = string("op_7117_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7119_cast_fp16 = pow(x = x_227_cast_fp16, y = var_7117_promoted_to_fp16)[name = string("op_7119_cast_fp16")];
            tensor<int32, [1]> var_7121_axes_0 = const()[name = string("op_7121_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7121_keep_dims_0 = const()[name = string("op_7121_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7121_cast_fp16 = reduce_mean(axes = var_7121_axes_0, keep_dims = var_7121_keep_dims_0, x = var_7119_cast_fp16)[name = string("op_7121_cast_fp16")];
            fp16 var_7122_to_fp16 = const()[name = string("op_7122_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7123_cast_fp16 = add(x = var_7121_cast_fp16, y = var_7122_to_fp16)[name = string("op_7123_cast_fp16")];
            fp32 var_7124_epsilon_0 = const()[name = string("op_7124_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7124_cast_fp16 = rsqrt(epsilon = var_7124_epsilon_0, x = var_7123_cast_fp16)[name = string("op_7124_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7125_cast_fp16 = mul(x = x_227_cast_fp16, y = var_7124_cast_fp16)[name = string("op_7125_cast_fp16")];
            tensor<fp16, [1024]> model_layers_18_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_18_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601215168)))];
            tensor<fp16, [1, 16, 1024]> input_185_cast_fp16 = mul(x = var_7125_cast_fp16, y = model_layers_18_post_attention_layernorm_weight_to_fp16)[name = string("input_185_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(601217280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604363072))))[name = string("model_layers_18_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_130_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_gate_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_130_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7130_cast_fp16 = silu(x = linear_130_cast_fp16)[name = string("op_7130_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_18_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(604369280))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607515072))))[name = string("model_layers_18_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_131_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_18_mlp_up_proj_weight_to_fp16_quantized, x = input_185_cast_fp16)[name = string("linear_131_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_189_cast_fp16 = mul(x = var_7130_cast_fp16, y = linear_131_cast_fp16)[name = string("input_189_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_18_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(607521280))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610667072))))[name = string("model_layers_18_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_132_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_18_mlp_down_proj_weight_to_fp16_quantized, x = input_189_cast_fp16)[name = string("linear_132_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_229_cast_fp16 = add(x = x_227_cast_fp16, y = linear_132_cast_fp16)[name = string("x_229_cast_fp16")];
            fp16 var_7144_promoted_to_fp16 = const()[name = string("op_7144_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7146_cast_fp16 = pow(x = x_229_cast_fp16, y = var_7144_promoted_to_fp16)[name = string("op_7146_cast_fp16")];
            tensor<int32, [1]> var_7148_axes_0 = const()[name = string("op_7148_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7148_keep_dims_0 = const()[name = string("op_7148_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7148_cast_fp16 = reduce_mean(axes = var_7148_axes_0, keep_dims = var_7148_keep_dims_0, x = var_7146_cast_fp16)[name = string("op_7148_cast_fp16")];
            fp16 var_7149_to_fp16 = const()[name = string("op_7149_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7150_cast_fp16 = add(x = var_7148_cast_fp16, y = var_7149_to_fp16)[name = string("op_7150_cast_fp16")];
            fp32 var_7151_epsilon_0 = const()[name = string("op_7151_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7151_cast_fp16 = rsqrt(epsilon = var_7151_epsilon_0, x = var_7150_cast_fp16)[name = string("op_7151_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7152_cast_fp16 = mul(x = x_229_cast_fp16, y = var_7151_cast_fp16)[name = string("op_7152_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610669184)))];
            tensor<fp16, [1, 16, 1024]> input_191_cast_fp16 = mul(x = var_7152_cast_fp16, y = model_layers_19_input_layernorm_weight_to_fp16)[name = string("input_191_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_19_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(610671296))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612768512))))[name = string("model_layers_19_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_133_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_19_self_attn_q_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_133_cast_fp16")];
            tensor<int32, [4]> var_7161 = const()[name = string("op_7161"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7162_cast_fp16 = reshape(shape = var_7161, x = linear_133_cast_fp16)[name = string("op_7162_cast_fp16")];
            tensor<int32, [4]> x_231_perm_0 = const()[name = string("x_231_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7170_promoted_to_fp16 = const()[name = string("op_7170_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_231_cast_fp16 = transpose(perm = x_231_perm_0, x = var_7162_cast_fp16)[name = string("transpose_44")];
            tensor<fp16, [1, 16, 16, 128]> var_7172_cast_fp16 = pow(x = x_231_cast_fp16, y = var_7170_promoted_to_fp16)[name = string("op_7172_cast_fp16")];
            tensor<int32, [1]> var_7174_axes_0 = const()[name = string("op_7174_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7174_keep_dims_0 = const()[name = string("op_7174_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7174_cast_fp16 = reduce_mean(axes = var_7174_axes_0, keep_dims = var_7174_keep_dims_0, x = var_7172_cast_fp16)[name = string("op_7174_cast_fp16")];
            fp16 var_7175_to_fp16 = const()[name = string("op_7175_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7176_cast_fp16 = add(x = var_7174_cast_fp16, y = var_7175_to_fp16)[name = string("op_7176_cast_fp16")];
            fp32 var_7177_epsilon_0 = const()[name = string("op_7177_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7177_cast_fp16 = rsqrt(epsilon = var_7177_epsilon_0, x = var_7176_cast_fp16)[name = string("op_7177_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7178_cast_fp16 = mul(x = x_231_cast_fp16, y = var_7177_cast_fp16)[name = string("op_7178_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772672)))];
            tensor<fp16, [1, 16, 16, 128]> x_235_cast_fp16 = mul(x = var_7178_cast_fp16, y = model_layers_19_self_attn_q_norm_weight_to_fp16)[name = string("x_235_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(612772992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613821632))))[name = string("model_layers_19_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_134_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_k_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_134_cast_fp16")];
            tensor<int32, [4]> var_7187 = const()[name = string("op_7187"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7188_cast_fp16 = reshape(shape = var_7187, x = linear_134_cast_fp16)[name = string("op_7188_cast_fp16")];
            tensor<int32, [4]> x_233_perm_0 = const()[name = string("x_233_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7196_promoted_to_fp16 = const()[name = string("op_7196_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_233_cast_fp16 = transpose(perm = x_233_perm_0, x = var_7188_cast_fp16)[name = string("transpose_43")];
            tensor<fp16, [1, 8, 16, 128]> var_7198_cast_fp16 = pow(x = x_233_cast_fp16, y = var_7196_promoted_to_fp16)[name = string("op_7198_cast_fp16")];
            tensor<int32, [1]> var_7200_axes_0 = const()[name = string("op_7200_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7200_keep_dims_0 = const()[name = string("op_7200_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7200_cast_fp16 = reduce_mean(axes = var_7200_axes_0, keep_dims = var_7200_keep_dims_0, x = var_7198_cast_fp16)[name = string("op_7200_cast_fp16")];
            fp16 var_7201_to_fp16 = const()[name = string("op_7201_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7202_cast_fp16 = add(x = var_7200_cast_fp16, y = var_7201_to_fp16)[name = string("op_7202_cast_fp16")];
            fp32 var_7203_epsilon_0 = const()[name = string("op_7203_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7203_cast_fp16 = rsqrt(epsilon = var_7203_epsilon_0, x = var_7202_cast_fp16)[name = string("op_7203_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7204_cast_fp16 = mul(x = x_233_cast_fp16, y = var_7203_cast_fp16)[name = string("op_7204_cast_fp16")];
            tensor<fp16, [128]> model_layers_19_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_19_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613823744)))];
            tensor<fp16, [1, 8, 16, 128]> x_237_cast_fp16 = mul(x = var_7204_cast_fp16, y = model_layers_19_self_attn_k_norm_weight_to_fp16)[name = string("x_237_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_19_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(613824064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614872704))))[name = string("model_layers_19_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_135_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_v_proj_weight_to_fp16_quantized, x = input_191_cast_fp16)[name = string("linear_135_cast_fp16")];
            tensor<int32, [4]> var_7213 = const()[name = string("op_7213"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7214_cast_fp16 = reshape(shape = var_7213, x = linear_135_cast_fp16)[name = string("op_7214_cast_fp16")];
            tensor<int32, [4]> v_39_perm_0 = const()[name = string("v_39_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_77_begin_0 = const()[name = string("first_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_77_end_0 = const()[name = string("first_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_77_end_mask_0 = const()[name = string("first_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_77_cast_fp16 = slice_by_index(begin = first_77_begin_0, end = first_77_end_0, end_mask = first_77_end_mask_0, x = x_235_cast_fp16)[name = string("first_77_cast_fp16")];
            tensor<int32, [4]> second_77_begin_0 = const()[name = string("second_77_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_77_end_0 = const()[name = string("second_77_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_77_end_mask_0 = const()[name = string("second_77_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_77_cast_fp16 = slice_by_index(begin = second_77_begin_0, end = second_77_end_0, end_mask = second_77_end_mask_0, x = x_235_cast_fp16)[name = string("second_77_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7228_cast_fp16 = mul(x = first_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7228_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7229_cast_fp16 = mul(x = second_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7229_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7231_cast_fp16 = sub(x = var_7228_cast_fp16, y = var_7229_cast_fp16)[name = string("op_7231_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7232_cast_fp16 = mul(x = second_77_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7232_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7233_cast_fp16 = mul(x = first_77_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7233_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7235_cast_fp16 = add(x = var_7232_cast_fp16, y = var_7233_cast_fp16)[name = string("op_7235_cast_fp16")];
            int32 var_7237 = const()[name = string("op_7237"), val = int32(-1)];
            bool q_39_interleave_0 = const()[name = string("q_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_39_cast_fp16 = concat(axis = var_7237, interleave = q_39_interleave_0, values = (var_7231_cast_fp16, var_7235_cast_fp16))[name = string("q_39_cast_fp16")];
            tensor<int32, [4]> first_79_begin_0 = const()[name = string("first_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_79_end_0 = const()[name = string("first_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_79_end_mask_0 = const()[name = string("first_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_79_cast_fp16 = slice_by_index(begin = first_79_begin_0, end = first_79_end_0, end_mask = first_79_end_mask_0, x = x_237_cast_fp16)[name = string("first_79_cast_fp16")];
            tensor<int32, [4]> second_79_begin_0 = const()[name = string("second_79_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_79_end_0 = const()[name = string("second_79_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_79_end_mask_0 = const()[name = string("second_79_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_79_cast_fp16 = slice_by_index(begin = second_79_begin_0, end = second_79_end_0, end_mask = second_79_end_mask_0, x = x_237_cast_fp16)[name = string("second_79_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7249_cast_fp16 = mul(x = first_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7249_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7250_cast_fp16 = mul(x = second_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7250_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7252_cast_fp16 = sub(x = var_7249_cast_fp16, y = var_7250_cast_fp16)[name = string("op_7252_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7253_cast_fp16 = mul(x = second_79_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7253_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7254_cast_fp16 = mul(x = first_79_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7254_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7256_cast_fp16 = add(x = var_7253_cast_fp16, y = var_7254_cast_fp16)[name = string("op_7256_cast_fp16")];
            int32 var_7258 = const()[name = string("op_7258"), val = int32(-1)];
            bool k_39_interleave_0 = const()[name = string("k_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_39_cast_fp16 = concat(axis = var_7258, interleave = k_39_interleave_0, values = (var_7252_cast_fp16, var_7256_cast_fp16))[name = string("k_39_cast_fp16")];
            tensor<int32, [4]> var_7269_begin_0 = const()[name = string("op_7269_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7269_end_0 = const()[name = string("op_7269_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7269_end_mask_0 = const()[name = string("op_7269_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7269_stride_0 = const()[name = string("op_7269_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7269_cast_fp16 = slice_by_index(begin = var_7269_begin_0, end = var_7269_end_0, end_mask = var_7269_end_mask_0, stride = var_7269_stride_0, x = q_39_cast_fp16)[name = string("op_7269_cast_fp16")];
            tensor<int32, [4]> var_7279_begin_0 = const()[name = string("op_7279_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7279_end_0 = const()[name = string("op_7279_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7279_end_mask_0 = const()[name = string("op_7279_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7279_stride_0 = const()[name = string("op_7279_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7279_cast_fp16 = slice_by_index(begin = var_7279_begin_0, end = var_7279_end_0, end_mask = var_7279_end_mask_0, stride = var_7279_stride_0, x = q_39_cast_fp16)[name = string("op_7279_cast_fp16")];
            int32 var_7281 = const()[name = string("op_7281"), val = int32(2)];
            bool stacked_39_interleave_0 = const()[name = string("stacked_39_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_39_cast_fp16 = concat(axis = var_7281, interleave = stacked_39_interleave_0, values = (var_7269_cast_fp16, var_7279_cast_fp16))[name = string("stacked_39_cast_fp16")];
            tensor<int32, [4]> var_7287_begin_0 = const()[name = string("op_7287_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7287_end_0 = const()[name = string("op_7287_end_0"), val = tensor<int32, [4]>([20, 8, 528, 128])];
            tensor<bool, [4]> var_7287_end_mask_0 = const()[name = string("op_7287_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7287_cast_fp16 = slice_by_index(begin = var_7287_begin_0, end = var_7287_end_0, end_mask = var_7287_end_mask_0, x = read_state_0)[name = string("op_7287_cast_fp16")];
            tensor<int32, [4]> var_7297_begin_0 = const()[name = string("op_7297_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7297_end_0 = const()[name = string("op_7297_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7297_end_mask_0 = const()[name = string("op_7297_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7297_cast_fp16 = slice_by_index(begin = var_7297_begin_0, end = var_7297_end_0, end_mask = var_7297_end_mask_0, x = var_7287_cast_fp16)[name = string("op_7297_cast_fp16")];
            bool var_7306_transpose_x_1 = const()[name = string("op_7306_transpose_x_1"), val = bool(false)];
            bool var_7306_transpose_y_1 = const()[name = string("op_7306_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7306_cast_fp16 = matmul(transpose_x = var_7306_transpose_x_1, transpose_y = var_7306_transpose_y_1, x = var_7297_cast_fp16, y = stacked_39_cast_fp16)[name = string("op_7306_cast_fp16")];
            tensor<int32, [4]> var_7309_perm_0 = const()[name = string("op_7309_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7310_to_fp16 = const()[name = string("op_7310_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7309_cast_fp16 = transpose(perm = var_7309_perm_0, x = var_7306_cast_fp16)[name = string("transpose_42")];
            tensor<fp16, [1, 8, 32, 512]> var_7311_cast_fp16 = mul(x = var_7309_cast_fp16, y = var_7310_to_fp16)[name = string("op_7311_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_39_cast_fp16 = add(x = var_7311_cast_fp16, y = cacheMask)[name = string("cache_weights_39_cast_fp16")];
            bool var_7317_transpose_x_1 = const()[name = string("op_7317_transpose_x_1"), val = bool(false)];
            bool var_7317_transpose_y_1 = const()[name = string("op_7317_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7317_cast_fp16 = matmul(transpose_x = var_7317_transpose_x_1, transpose_y = var_7317_transpose_y_1, x = stacked_39_cast_fp16, y = k_39_cast_fp16)[name = string("op_7317_cast_fp16")];
            fp16 var_7318_to_fp16 = const()[name = string("op_7318_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7319_cast_fp16 = mul(x = var_7317_cast_fp16, y = var_7318_to_fp16)[name = string("op_7319_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_39_cast_fp16 = add(x = var_7319_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_39_cast_fp16")];
            int32 var_7323 = const()[name = string("op_7323"), val = int32(-1)];
            bool var_7324_interleave_0 = const()[name = string("op_7324_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7324_cast_fp16 = concat(axis = var_7323, interleave = var_7324_interleave_0, values = (cache_weights_39_cast_fp16, call_weights_39_cast_fp16))[name = string("op_7324_cast_fp16")];
            int32 var_7325 = const()[name = string("op_7325"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_39_cast_fp16 = softmax(axis = var_7325, x = var_7324_cast_fp16)[name = string("shares_39_cast_fp16")];
            tensor<int32, [4]> var_7332_begin_0 = const()[name = string("op_7332_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7332_end_0 = const()[name = string("op_7332_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7332_end_mask_0 = const()[name = string("op_7332_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7332_cast_fp16 = slice_by_index(begin = var_7332_begin_0, end = var_7332_end_0, end_mask = var_7332_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7332_cast_fp16")];
            tensor<int32, [4]> var_7337_begin_0 = const()[name = string("op_7337_begin_0"), val = tensor<int32, [4]>([19, 0, 0, 0])];
            tensor<int32, [4]> var_7337_end_0 = const()[name = string("op_7337_end_0"), val = tensor<int32, [4]>([20, 8, 528, 128])];
            tensor<bool, [4]> var_7337_end_mask_0 = const()[name = string("op_7337_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7337_cast_fp16 = slice_by_index(begin = var_7337_begin_0, end = var_7337_end_0, end_mask = var_7337_end_mask_0, x = read_state_1)[name = string("op_7337_cast_fp16")];
            tensor<int32, [4]> var_7347_begin_0 = const()[name = string("op_7347_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7347_end_0 = const()[name = string("op_7347_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7347_end_mask_0 = const()[name = string("op_7347_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7347_cast_fp16 = slice_by_index(begin = var_7347_begin_0, end = var_7347_end_0, end_mask = var_7347_end_mask_0, x = var_7337_cast_fp16)[name = string("op_7347_cast_fp16")];
            bool var_7353_transpose_x_0 = const()[name = string("op_7353_transpose_x_0"), val = bool(false)];
            bool var_7353_transpose_y_0 = const()[name = string("op_7353_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7353_cast_fp16 = matmul(transpose_x = var_7353_transpose_x_0, transpose_y = var_7353_transpose_y_0, x = var_7332_cast_fp16, y = var_7347_cast_fp16)[name = string("op_7353_cast_fp16")];
            tensor<int32, [4]> var_7358_begin_0 = const()[name = string("op_7358_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7358_end_0 = const()[name = string("op_7358_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7358_end_mask_0 = const()[name = string("op_7358_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7358_cast_fp16 = slice_by_index(begin = var_7358_begin_0, end = var_7358_end_0, end_mask = var_7358_end_mask_0, x = shares_39_cast_fp16)[name = string("op_7358_cast_fp16")];
            bool var_7359_transpose_x_0 = const()[name = string("op_7359_transpose_x_0"), val = bool(false)];
            bool var_7359_transpose_y_0 = const()[name = string("op_7359_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_39_cast_fp16 = transpose(perm = v_39_perm_0, x = var_7214_cast_fp16)[name = string("transpose_41")];
            tensor<fp16, [1, 8, 32, 128]> var_7359_cast_fp16 = matmul(transpose_x = var_7359_transpose_x_0, transpose_y = var_7359_transpose_y_0, x = var_7358_cast_fp16, y = v_39_cast_fp16)[name = string("op_7359_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_77_cast_fp16 = add(x = var_7353_cast_fp16, y = var_7359_cast_fp16)[name = string("mixed_77_cast_fp16")];
            tensor<int32, [4]> var_7376_begin_0 = const()[name = string("op_7376_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7376_end_0 = const()[name = string("op_7376_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7376_end_mask_0 = const()[name = string("op_7376_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7376_cast_fp16 = slice_by_index(begin = var_7376_begin_0, end = var_7376_end_0, end_mask = var_7376_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7376_cast_fp16")];
            tensor<int32, [4]> var_7391_begin_0 = const()[name = string("op_7391_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7391_end_0 = const()[name = string("op_7391_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7391_end_mask_0 = const()[name = string("op_7391_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7391_cast_fp16 = slice_by_index(begin = var_7391_begin_0, end = var_7391_end_0, end_mask = var_7391_end_mask_0, x = mixed_77_cast_fp16)[name = string("op_7391_cast_fp16")];
            int32 var_7393 = const()[name = string("op_7393"), val = int32(1)];
            bool mixed_79_interleave_0 = const()[name = string("mixed_79_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_79_cast_fp16 = concat(axis = var_7393, interleave = mixed_79_interleave_0, values = (var_7376_cast_fp16, var_7391_cast_fp16))[name = string("mixed_79_cast_fp16")];
            tensor<int32, [4]> var_7397_perm_0 = const()[name = string("op_7397_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7401 = const()[name = string("op_7401"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7397_cast_fp16 = transpose(perm = var_7397_perm_0, x = mixed_79_cast_fp16)[name = string("transpose_40")];
            tensor<fp16, [1, 16, 2048]> input_193_cast_fp16 = reshape(shape = var_7401, x = var_7397_cast_fp16)[name = string("input_193_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_19_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(614874816))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616972032))))[name = string("model_layers_19_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_136_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_self_attn_o_proj_weight_to_fp16_quantized, x = input_193_cast_fp16)[name = string("linear_136_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_239_cast_fp16 = add(x = x_229_cast_fp16, y = linear_136_cast_fp16)[name = string("x_239_cast_fp16")];
            fp16 var_7412_promoted_to_fp16 = const()[name = string("op_7412_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7414_cast_fp16 = pow(x = x_239_cast_fp16, y = var_7412_promoted_to_fp16)[name = string("op_7414_cast_fp16")];
            tensor<int32, [1]> var_7416_axes_0 = const()[name = string("op_7416_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7416_keep_dims_0 = const()[name = string("op_7416_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7416_cast_fp16 = reduce_mean(axes = var_7416_axes_0, keep_dims = var_7416_keep_dims_0, x = var_7414_cast_fp16)[name = string("op_7416_cast_fp16")];
            fp16 var_7417_to_fp16 = const()[name = string("op_7417_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7418_cast_fp16 = add(x = var_7416_cast_fp16, y = var_7417_to_fp16)[name = string("op_7418_cast_fp16")];
            fp32 var_7419_epsilon_0 = const()[name = string("op_7419_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7419_cast_fp16 = rsqrt(epsilon = var_7419_epsilon_0, x = var_7418_cast_fp16)[name = string("op_7419_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7420_cast_fp16 = mul(x = x_239_cast_fp16, y = var_7419_cast_fp16)[name = string("op_7420_cast_fp16")];
            tensor<fp16, [1024]> model_layers_19_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_19_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616974144)))];
            tensor<fp16, [1, 16, 1024]> input_195_cast_fp16 = mul(x = var_7420_cast_fp16, y = model_layers_19_post_attention_layernorm_weight_to_fp16)[name = string("input_195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(616976256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620122048))))[name = string("model_layers_19_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_137_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_gate_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_137_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7425_cast_fp16 = silu(x = linear_137_cast_fp16)[name = string("op_7425_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_19_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(620128256))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623274048))))[name = string("model_layers_19_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_138_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_19_mlp_up_proj_weight_to_fp16_quantized, x = input_195_cast_fp16)[name = string("linear_138_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_199_cast_fp16 = mul(x = var_7425_cast_fp16, y = linear_138_cast_fp16)[name = string("input_199_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_19_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(623280256))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626426048))))[name = string("model_layers_19_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_139_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_19_mlp_down_proj_weight_to_fp16_quantized, x = input_199_cast_fp16)[name = string("linear_139_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_241_cast_fp16 = add(x = x_239_cast_fp16, y = linear_139_cast_fp16)[name = string("x_241_cast_fp16")];
            fp16 var_7439_promoted_to_fp16 = const()[name = string("op_7439_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7441_cast_fp16 = pow(x = x_241_cast_fp16, y = var_7439_promoted_to_fp16)[name = string("op_7441_cast_fp16")];
            tensor<int32, [1]> var_7443_axes_0 = const()[name = string("op_7443_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7443_keep_dims_0 = const()[name = string("op_7443_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7443_cast_fp16 = reduce_mean(axes = var_7443_axes_0, keep_dims = var_7443_keep_dims_0, x = var_7441_cast_fp16)[name = string("op_7443_cast_fp16")];
            fp16 var_7444_to_fp16 = const()[name = string("op_7444_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7445_cast_fp16 = add(x = var_7443_cast_fp16, y = var_7444_to_fp16)[name = string("op_7445_cast_fp16")];
            fp32 var_7446_epsilon_0 = const()[name = string("op_7446_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7446_cast_fp16 = rsqrt(epsilon = var_7446_epsilon_0, x = var_7445_cast_fp16)[name = string("op_7446_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7447_cast_fp16 = mul(x = x_241_cast_fp16, y = var_7446_cast_fp16)[name = string("op_7447_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626428160)))];
            tensor<fp16, [1, 16, 1024]> input_201_cast_fp16 = mul(x = var_7447_cast_fp16, y = model_layers_20_input_layernorm_weight_to_fp16)[name = string("input_201_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_20_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(626430272))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628527488))))[name = string("model_layers_20_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_140_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_20_self_attn_q_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_140_cast_fp16")];
            tensor<int32, [4]> var_7456 = const()[name = string("op_7456"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7457_cast_fp16 = reshape(shape = var_7456, x = linear_140_cast_fp16)[name = string("op_7457_cast_fp16")];
            tensor<int32, [4]> x_243_perm_0 = const()[name = string("x_243_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7465_promoted_to_fp16 = const()[name = string("op_7465_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_243_cast_fp16 = transpose(perm = x_243_perm_0, x = var_7457_cast_fp16)[name = string("transpose_39")];
            tensor<fp16, [1, 16, 16, 128]> var_7467_cast_fp16 = pow(x = x_243_cast_fp16, y = var_7465_promoted_to_fp16)[name = string("op_7467_cast_fp16")];
            tensor<int32, [1]> var_7469_axes_0 = const()[name = string("op_7469_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7469_keep_dims_0 = const()[name = string("op_7469_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7469_cast_fp16 = reduce_mean(axes = var_7469_axes_0, keep_dims = var_7469_keep_dims_0, x = var_7467_cast_fp16)[name = string("op_7469_cast_fp16")];
            fp16 var_7470_to_fp16 = const()[name = string("op_7470_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7471_cast_fp16 = add(x = var_7469_cast_fp16, y = var_7470_to_fp16)[name = string("op_7471_cast_fp16")];
            fp32 var_7472_epsilon_0 = const()[name = string("op_7472_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7472_cast_fp16 = rsqrt(epsilon = var_7472_epsilon_0, x = var_7471_cast_fp16)[name = string("op_7472_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7473_cast_fp16 = mul(x = x_243_cast_fp16, y = var_7472_cast_fp16)[name = string("op_7473_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531648)))];
            tensor<fp16, [1, 16, 16, 128]> x_247_cast_fp16 = mul(x = var_7473_cast_fp16, y = model_layers_20_self_attn_q_norm_weight_to_fp16)[name = string("x_247_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(628531968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629580608))))[name = string("model_layers_20_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_141_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_k_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_141_cast_fp16")];
            tensor<int32, [4]> var_7482 = const()[name = string("op_7482"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7483_cast_fp16 = reshape(shape = var_7482, x = linear_141_cast_fp16)[name = string("op_7483_cast_fp16")];
            tensor<int32, [4]> x_245_perm_0 = const()[name = string("x_245_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7491_promoted_to_fp16 = const()[name = string("op_7491_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_245_cast_fp16 = transpose(perm = x_245_perm_0, x = var_7483_cast_fp16)[name = string("transpose_38")];
            tensor<fp16, [1, 8, 16, 128]> var_7493_cast_fp16 = pow(x = x_245_cast_fp16, y = var_7491_promoted_to_fp16)[name = string("op_7493_cast_fp16")];
            tensor<int32, [1]> var_7495_axes_0 = const()[name = string("op_7495_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7495_keep_dims_0 = const()[name = string("op_7495_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7495_cast_fp16 = reduce_mean(axes = var_7495_axes_0, keep_dims = var_7495_keep_dims_0, x = var_7493_cast_fp16)[name = string("op_7495_cast_fp16")];
            fp16 var_7496_to_fp16 = const()[name = string("op_7496_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7497_cast_fp16 = add(x = var_7495_cast_fp16, y = var_7496_to_fp16)[name = string("op_7497_cast_fp16")];
            fp32 var_7498_epsilon_0 = const()[name = string("op_7498_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7498_cast_fp16 = rsqrt(epsilon = var_7498_epsilon_0, x = var_7497_cast_fp16)[name = string("op_7498_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7499_cast_fp16 = mul(x = x_245_cast_fp16, y = var_7498_cast_fp16)[name = string("op_7499_cast_fp16")];
            tensor<fp16, [128]> model_layers_20_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_20_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629582720)))];
            tensor<fp16, [1, 8, 16, 128]> x_249_cast_fp16 = mul(x = var_7499_cast_fp16, y = model_layers_20_self_attn_k_norm_weight_to_fp16)[name = string("x_249_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_20_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(629583040))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630631680))))[name = string("model_layers_20_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_142_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_v_proj_weight_to_fp16_quantized, x = input_201_cast_fp16)[name = string("linear_142_cast_fp16")];
            tensor<int32, [4]> var_7508 = const()[name = string("op_7508"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7509_cast_fp16 = reshape(shape = var_7508, x = linear_142_cast_fp16)[name = string("op_7509_cast_fp16")];
            tensor<int32, [4]> v_41_perm_0 = const()[name = string("v_41_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_81_begin_0 = const()[name = string("first_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_81_end_0 = const()[name = string("first_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_81_end_mask_0 = const()[name = string("first_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_81_cast_fp16 = slice_by_index(begin = first_81_begin_0, end = first_81_end_0, end_mask = first_81_end_mask_0, x = x_247_cast_fp16)[name = string("first_81_cast_fp16")];
            tensor<int32, [4]> second_81_begin_0 = const()[name = string("second_81_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_81_end_0 = const()[name = string("second_81_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_81_end_mask_0 = const()[name = string("second_81_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_81_cast_fp16 = slice_by_index(begin = second_81_begin_0, end = second_81_end_0, end_mask = second_81_end_mask_0, x = x_247_cast_fp16)[name = string("second_81_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7523_cast_fp16 = mul(x = first_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7523_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7524_cast_fp16 = mul(x = second_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7524_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7526_cast_fp16 = sub(x = var_7523_cast_fp16, y = var_7524_cast_fp16)[name = string("op_7526_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7527_cast_fp16 = mul(x = second_81_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7527_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7528_cast_fp16 = mul(x = first_81_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7528_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7530_cast_fp16 = add(x = var_7527_cast_fp16, y = var_7528_cast_fp16)[name = string("op_7530_cast_fp16")];
            int32 var_7532 = const()[name = string("op_7532"), val = int32(-1)];
            bool q_41_interleave_0 = const()[name = string("q_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_41_cast_fp16 = concat(axis = var_7532, interleave = q_41_interleave_0, values = (var_7526_cast_fp16, var_7530_cast_fp16))[name = string("q_41_cast_fp16")];
            tensor<int32, [4]> first_83_begin_0 = const()[name = string("first_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_83_end_0 = const()[name = string("first_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_83_end_mask_0 = const()[name = string("first_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_83_cast_fp16 = slice_by_index(begin = first_83_begin_0, end = first_83_end_0, end_mask = first_83_end_mask_0, x = x_249_cast_fp16)[name = string("first_83_cast_fp16")];
            tensor<int32, [4]> second_83_begin_0 = const()[name = string("second_83_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_83_end_0 = const()[name = string("second_83_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_83_end_mask_0 = const()[name = string("second_83_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_83_cast_fp16 = slice_by_index(begin = second_83_begin_0, end = second_83_end_0, end_mask = second_83_end_mask_0, x = x_249_cast_fp16)[name = string("second_83_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7544_cast_fp16 = mul(x = first_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7544_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7545_cast_fp16 = mul(x = second_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7545_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7547_cast_fp16 = sub(x = var_7544_cast_fp16, y = var_7545_cast_fp16)[name = string("op_7547_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7548_cast_fp16 = mul(x = second_83_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7548_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7549_cast_fp16 = mul(x = first_83_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7549_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7551_cast_fp16 = add(x = var_7548_cast_fp16, y = var_7549_cast_fp16)[name = string("op_7551_cast_fp16")];
            int32 var_7553 = const()[name = string("op_7553"), val = int32(-1)];
            bool k_41_interleave_0 = const()[name = string("k_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_41_cast_fp16 = concat(axis = var_7553, interleave = k_41_interleave_0, values = (var_7547_cast_fp16, var_7551_cast_fp16))[name = string("k_41_cast_fp16")];
            tensor<int32, [4]> var_7564_begin_0 = const()[name = string("op_7564_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7564_end_0 = const()[name = string("op_7564_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7564_end_mask_0 = const()[name = string("op_7564_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7564_stride_0 = const()[name = string("op_7564_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7564_cast_fp16 = slice_by_index(begin = var_7564_begin_0, end = var_7564_end_0, end_mask = var_7564_end_mask_0, stride = var_7564_stride_0, x = q_41_cast_fp16)[name = string("op_7564_cast_fp16")];
            tensor<int32, [4]> var_7574_begin_0 = const()[name = string("op_7574_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7574_end_0 = const()[name = string("op_7574_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7574_end_mask_0 = const()[name = string("op_7574_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7574_stride_0 = const()[name = string("op_7574_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7574_cast_fp16 = slice_by_index(begin = var_7574_begin_0, end = var_7574_end_0, end_mask = var_7574_end_mask_0, stride = var_7574_stride_0, x = q_41_cast_fp16)[name = string("op_7574_cast_fp16")];
            int32 var_7576 = const()[name = string("op_7576"), val = int32(2)];
            bool stacked_41_interleave_0 = const()[name = string("stacked_41_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_41_cast_fp16 = concat(axis = var_7576, interleave = stacked_41_interleave_0, values = (var_7564_cast_fp16, var_7574_cast_fp16))[name = string("stacked_41_cast_fp16")];
            tensor<int32, [4]> var_7582_begin_0 = const()[name = string("op_7582_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7582_end_0 = const()[name = string("op_7582_end_0"), val = tensor<int32, [4]>([21, 8, 528, 128])];
            tensor<bool, [4]> var_7582_end_mask_0 = const()[name = string("op_7582_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7582_cast_fp16 = slice_by_index(begin = var_7582_begin_0, end = var_7582_end_0, end_mask = var_7582_end_mask_0, x = read_state_0)[name = string("op_7582_cast_fp16")];
            tensor<int32, [4]> var_7592_begin_0 = const()[name = string("op_7592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7592_end_0 = const()[name = string("op_7592_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7592_end_mask_0 = const()[name = string("op_7592_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7592_cast_fp16 = slice_by_index(begin = var_7592_begin_0, end = var_7592_end_0, end_mask = var_7592_end_mask_0, x = var_7582_cast_fp16)[name = string("op_7592_cast_fp16")];
            bool var_7601_transpose_x_1 = const()[name = string("op_7601_transpose_x_1"), val = bool(false)];
            bool var_7601_transpose_y_1 = const()[name = string("op_7601_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7601_cast_fp16 = matmul(transpose_x = var_7601_transpose_x_1, transpose_y = var_7601_transpose_y_1, x = var_7592_cast_fp16, y = stacked_41_cast_fp16)[name = string("op_7601_cast_fp16")];
            tensor<int32, [4]> var_7604_perm_0 = const()[name = string("op_7604_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7605_to_fp16 = const()[name = string("op_7605_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7604_cast_fp16 = transpose(perm = var_7604_perm_0, x = var_7601_cast_fp16)[name = string("transpose_37")];
            tensor<fp16, [1, 8, 32, 512]> var_7606_cast_fp16 = mul(x = var_7604_cast_fp16, y = var_7605_to_fp16)[name = string("op_7606_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_41_cast_fp16 = add(x = var_7606_cast_fp16, y = cacheMask)[name = string("cache_weights_41_cast_fp16")];
            bool var_7612_transpose_x_1 = const()[name = string("op_7612_transpose_x_1"), val = bool(false)];
            bool var_7612_transpose_y_1 = const()[name = string("op_7612_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7612_cast_fp16 = matmul(transpose_x = var_7612_transpose_x_1, transpose_y = var_7612_transpose_y_1, x = stacked_41_cast_fp16, y = k_41_cast_fp16)[name = string("op_7612_cast_fp16")];
            fp16 var_7613_to_fp16 = const()[name = string("op_7613_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7614_cast_fp16 = mul(x = var_7612_cast_fp16, y = var_7613_to_fp16)[name = string("op_7614_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_41_cast_fp16 = add(x = var_7614_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_41_cast_fp16")];
            int32 var_7618 = const()[name = string("op_7618"), val = int32(-1)];
            bool var_7619_interleave_0 = const()[name = string("op_7619_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7619_cast_fp16 = concat(axis = var_7618, interleave = var_7619_interleave_0, values = (cache_weights_41_cast_fp16, call_weights_41_cast_fp16))[name = string("op_7619_cast_fp16")];
            int32 var_7620 = const()[name = string("op_7620"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_41_cast_fp16 = softmax(axis = var_7620, x = var_7619_cast_fp16)[name = string("shares_41_cast_fp16")];
            tensor<int32, [4]> var_7627_begin_0 = const()[name = string("op_7627_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7627_end_0 = const()[name = string("op_7627_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7627_end_mask_0 = const()[name = string("op_7627_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7627_cast_fp16 = slice_by_index(begin = var_7627_begin_0, end = var_7627_end_0, end_mask = var_7627_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7627_cast_fp16")];
            tensor<int32, [4]> var_7632_begin_0 = const()[name = string("op_7632_begin_0"), val = tensor<int32, [4]>([20, 0, 0, 0])];
            tensor<int32, [4]> var_7632_end_0 = const()[name = string("op_7632_end_0"), val = tensor<int32, [4]>([21, 8, 528, 128])];
            tensor<bool, [4]> var_7632_end_mask_0 = const()[name = string("op_7632_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7632_cast_fp16 = slice_by_index(begin = var_7632_begin_0, end = var_7632_end_0, end_mask = var_7632_end_mask_0, x = read_state_1)[name = string("op_7632_cast_fp16")];
            tensor<int32, [4]> var_7642_begin_0 = const()[name = string("op_7642_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7642_end_0 = const()[name = string("op_7642_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7642_end_mask_0 = const()[name = string("op_7642_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7642_cast_fp16 = slice_by_index(begin = var_7642_begin_0, end = var_7642_end_0, end_mask = var_7642_end_mask_0, x = var_7632_cast_fp16)[name = string("op_7642_cast_fp16")];
            bool var_7648_transpose_x_0 = const()[name = string("op_7648_transpose_x_0"), val = bool(false)];
            bool var_7648_transpose_y_0 = const()[name = string("op_7648_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7648_cast_fp16 = matmul(transpose_x = var_7648_transpose_x_0, transpose_y = var_7648_transpose_y_0, x = var_7627_cast_fp16, y = var_7642_cast_fp16)[name = string("op_7648_cast_fp16")];
            tensor<int32, [4]> var_7653_begin_0 = const()[name = string("op_7653_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7653_end_0 = const()[name = string("op_7653_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7653_end_mask_0 = const()[name = string("op_7653_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7653_cast_fp16 = slice_by_index(begin = var_7653_begin_0, end = var_7653_end_0, end_mask = var_7653_end_mask_0, x = shares_41_cast_fp16)[name = string("op_7653_cast_fp16")];
            bool var_7654_transpose_x_0 = const()[name = string("op_7654_transpose_x_0"), val = bool(false)];
            bool var_7654_transpose_y_0 = const()[name = string("op_7654_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_41_cast_fp16 = transpose(perm = v_41_perm_0, x = var_7509_cast_fp16)[name = string("transpose_36")];
            tensor<fp16, [1, 8, 32, 128]> var_7654_cast_fp16 = matmul(transpose_x = var_7654_transpose_x_0, transpose_y = var_7654_transpose_y_0, x = var_7653_cast_fp16, y = v_41_cast_fp16)[name = string("op_7654_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_81_cast_fp16 = add(x = var_7648_cast_fp16, y = var_7654_cast_fp16)[name = string("mixed_81_cast_fp16")];
            tensor<int32, [4]> var_7671_begin_0 = const()[name = string("op_7671_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7671_end_0 = const()[name = string("op_7671_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7671_end_mask_0 = const()[name = string("op_7671_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7671_cast_fp16 = slice_by_index(begin = var_7671_begin_0, end = var_7671_end_0, end_mask = var_7671_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7671_cast_fp16")];
            tensor<int32, [4]> var_7686_begin_0 = const()[name = string("op_7686_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7686_end_0 = const()[name = string("op_7686_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7686_end_mask_0 = const()[name = string("op_7686_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7686_cast_fp16 = slice_by_index(begin = var_7686_begin_0, end = var_7686_end_0, end_mask = var_7686_end_mask_0, x = mixed_81_cast_fp16)[name = string("op_7686_cast_fp16")];
            int32 var_7688 = const()[name = string("op_7688"), val = int32(1)];
            bool mixed_83_interleave_0 = const()[name = string("mixed_83_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_83_cast_fp16 = concat(axis = var_7688, interleave = mixed_83_interleave_0, values = (var_7671_cast_fp16, var_7686_cast_fp16))[name = string("mixed_83_cast_fp16")];
            tensor<int32, [4]> var_7692_perm_0 = const()[name = string("op_7692_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7696 = const()[name = string("op_7696"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7692_cast_fp16 = transpose(perm = var_7692_perm_0, x = mixed_83_cast_fp16)[name = string("transpose_35")];
            tensor<fp16, [1, 16, 2048]> input_203_cast_fp16 = reshape(shape = var_7696, x = var_7692_cast_fp16)[name = string("input_203_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_20_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(630633792))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632731008))))[name = string("model_layers_20_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_143_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_self_attn_o_proj_weight_to_fp16_quantized, x = input_203_cast_fp16)[name = string("linear_143_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_251_cast_fp16 = add(x = x_241_cast_fp16, y = linear_143_cast_fp16)[name = string("x_251_cast_fp16")];
            fp16 var_7707_promoted_to_fp16 = const()[name = string("op_7707_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7709_cast_fp16 = pow(x = x_251_cast_fp16, y = var_7707_promoted_to_fp16)[name = string("op_7709_cast_fp16")];
            tensor<int32, [1]> var_7711_axes_0 = const()[name = string("op_7711_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7711_keep_dims_0 = const()[name = string("op_7711_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7711_cast_fp16 = reduce_mean(axes = var_7711_axes_0, keep_dims = var_7711_keep_dims_0, x = var_7709_cast_fp16)[name = string("op_7711_cast_fp16")];
            fp16 var_7712_to_fp16 = const()[name = string("op_7712_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7713_cast_fp16 = add(x = var_7711_cast_fp16, y = var_7712_to_fp16)[name = string("op_7713_cast_fp16")];
            fp32 var_7714_epsilon_0 = const()[name = string("op_7714_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7714_cast_fp16 = rsqrt(epsilon = var_7714_epsilon_0, x = var_7713_cast_fp16)[name = string("op_7714_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7715_cast_fp16 = mul(x = x_251_cast_fp16, y = var_7714_cast_fp16)[name = string("op_7715_cast_fp16")];
            tensor<fp16, [1024]> model_layers_20_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_20_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632733120)))];
            tensor<fp16, [1, 16, 1024]> input_205_cast_fp16 = mul(x = var_7715_cast_fp16, y = model_layers_20_post_attention_layernorm_weight_to_fp16)[name = string("input_205_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(632735232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635881024))))[name = string("model_layers_20_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_144_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_gate_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_144_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_7720_cast_fp16 = silu(x = linear_144_cast_fp16)[name = string("op_7720_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_20_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(635887232))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639033024))))[name = string("model_layers_20_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_145_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_20_mlp_up_proj_weight_to_fp16_quantized, x = input_205_cast_fp16)[name = string("linear_145_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_209_cast_fp16 = mul(x = var_7720_cast_fp16, y = linear_145_cast_fp16)[name = string("input_209_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_20_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(639039232))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642185024))))[name = string("model_layers_20_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_146_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_20_mlp_down_proj_weight_to_fp16_quantized, x = input_209_cast_fp16)[name = string("linear_146_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_253_cast_fp16 = add(x = x_251_cast_fp16, y = linear_146_cast_fp16)[name = string("x_253_cast_fp16")];
            fp16 var_7734_promoted_to_fp16 = const()[name = string("op_7734_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_7736_cast_fp16 = pow(x = x_253_cast_fp16, y = var_7734_promoted_to_fp16)[name = string("op_7736_cast_fp16")];
            tensor<int32, [1]> var_7738_axes_0 = const()[name = string("op_7738_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7738_keep_dims_0 = const()[name = string("op_7738_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_7738_cast_fp16 = reduce_mean(axes = var_7738_axes_0, keep_dims = var_7738_keep_dims_0, x = var_7736_cast_fp16)[name = string("op_7738_cast_fp16")];
            fp16 var_7739_to_fp16 = const()[name = string("op_7739_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_7740_cast_fp16 = add(x = var_7738_cast_fp16, y = var_7739_to_fp16)[name = string("op_7740_cast_fp16")];
            fp32 var_7741_epsilon_0 = const()[name = string("op_7741_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_7741_cast_fp16 = rsqrt(epsilon = var_7741_epsilon_0, x = var_7740_cast_fp16)[name = string("op_7741_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_7742_cast_fp16 = mul(x = x_253_cast_fp16, y = var_7741_cast_fp16)[name = string("op_7742_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642187136)))];
            tensor<fp16, [1, 16, 1024]> input_211_cast_fp16 = mul(x = var_7742_cast_fp16, y = model_layers_21_input_layernorm_weight_to_fp16)[name = string("input_211_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_21_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(642189248))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644286464))))[name = string("model_layers_21_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_147_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_21_self_attn_q_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_147_cast_fp16")];
            tensor<int32, [4]> var_7751 = const()[name = string("op_7751"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_7752_cast_fp16 = reshape(shape = var_7751, x = linear_147_cast_fp16)[name = string("op_7752_cast_fp16")];
            tensor<int32, [4]> x_255_perm_0 = const()[name = string("x_255_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7760_promoted_to_fp16 = const()[name = string("op_7760_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_255_cast_fp16 = transpose(perm = x_255_perm_0, x = var_7752_cast_fp16)[name = string("transpose_34")];
            tensor<fp16, [1, 16, 16, 128]> var_7762_cast_fp16 = pow(x = x_255_cast_fp16, y = var_7760_promoted_to_fp16)[name = string("op_7762_cast_fp16")];
            tensor<int32, [1]> var_7764_axes_0 = const()[name = string("op_7764_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7764_keep_dims_0 = const()[name = string("op_7764_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_7764_cast_fp16 = reduce_mean(axes = var_7764_axes_0, keep_dims = var_7764_keep_dims_0, x = var_7762_cast_fp16)[name = string("op_7764_cast_fp16")];
            fp16 var_7765_to_fp16 = const()[name = string("op_7765_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_7766_cast_fp16 = add(x = var_7764_cast_fp16, y = var_7765_to_fp16)[name = string("op_7766_cast_fp16")];
            fp32 var_7767_epsilon_0 = const()[name = string("op_7767_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_7767_cast_fp16 = rsqrt(epsilon = var_7767_epsilon_0, x = var_7766_cast_fp16)[name = string("op_7767_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_7768_cast_fp16 = mul(x = x_255_cast_fp16, y = var_7767_cast_fp16)[name = string("op_7768_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290624)))];
            tensor<fp16, [1, 16, 16, 128]> x_259_cast_fp16 = mul(x = var_7768_cast_fp16, y = model_layers_21_self_attn_q_norm_weight_to_fp16)[name = string("x_259_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(644290944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645339584))))[name = string("model_layers_21_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_148_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_k_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_148_cast_fp16")];
            tensor<int32, [4]> var_7777 = const()[name = string("op_7777"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7778_cast_fp16 = reshape(shape = var_7777, x = linear_148_cast_fp16)[name = string("op_7778_cast_fp16")];
            tensor<int32, [4]> x_257_perm_0 = const()[name = string("x_257_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_7786_promoted_to_fp16 = const()[name = string("op_7786_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_257_cast_fp16 = transpose(perm = x_257_perm_0, x = var_7778_cast_fp16)[name = string("transpose_33")];
            tensor<fp16, [1, 8, 16, 128]> var_7788_cast_fp16 = pow(x = x_257_cast_fp16, y = var_7786_promoted_to_fp16)[name = string("op_7788_cast_fp16")];
            tensor<int32, [1]> var_7790_axes_0 = const()[name = string("op_7790_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_7790_keep_dims_0 = const()[name = string("op_7790_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_7790_cast_fp16 = reduce_mean(axes = var_7790_axes_0, keep_dims = var_7790_keep_dims_0, x = var_7788_cast_fp16)[name = string("op_7790_cast_fp16")];
            fp16 var_7791_to_fp16 = const()[name = string("op_7791_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_7792_cast_fp16 = add(x = var_7790_cast_fp16, y = var_7791_to_fp16)[name = string("op_7792_cast_fp16")];
            fp32 var_7793_epsilon_0 = const()[name = string("op_7793_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_7793_cast_fp16 = rsqrt(epsilon = var_7793_epsilon_0, x = var_7792_cast_fp16)[name = string("op_7793_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_7794_cast_fp16 = mul(x = x_257_cast_fp16, y = var_7793_cast_fp16)[name = string("op_7794_cast_fp16")];
            tensor<fp16, [128]> model_layers_21_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_21_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645341696)))];
            tensor<fp16, [1, 8, 16, 128]> x_261_cast_fp16 = mul(x = var_7794_cast_fp16, y = model_layers_21_self_attn_k_norm_weight_to_fp16)[name = string("x_261_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_21_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(645342016))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646390656))))[name = string("model_layers_21_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_149_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_v_proj_weight_to_fp16_quantized, x = input_211_cast_fp16)[name = string("linear_149_cast_fp16")];
            tensor<int32, [4]> var_7803 = const()[name = string("op_7803"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_7804_cast_fp16 = reshape(shape = var_7803, x = linear_149_cast_fp16)[name = string("op_7804_cast_fp16")];
            tensor<int32, [4]> v_43_perm_0 = const()[name = string("v_43_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_85_begin_0 = const()[name = string("first_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_85_end_0 = const()[name = string("first_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_85_end_mask_0 = const()[name = string("first_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_85_cast_fp16 = slice_by_index(begin = first_85_begin_0, end = first_85_end_0, end_mask = first_85_end_mask_0, x = x_259_cast_fp16)[name = string("first_85_cast_fp16")];
            tensor<int32, [4]> second_85_begin_0 = const()[name = string("second_85_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_85_end_0 = const()[name = string("second_85_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_85_end_mask_0 = const()[name = string("second_85_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_85_cast_fp16 = slice_by_index(begin = second_85_begin_0, end = second_85_end_0, end_mask = second_85_end_mask_0, x = x_259_cast_fp16)[name = string("second_85_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7818_cast_fp16 = mul(x = first_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7818_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7819_cast_fp16 = mul(x = second_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7819_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7821_cast_fp16 = sub(x = var_7818_cast_fp16, y = var_7819_cast_fp16)[name = string("op_7821_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7822_cast_fp16 = mul(x = second_85_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7822_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7823_cast_fp16 = mul(x = first_85_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7823_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_7825_cast_fp16 = add(x = var_7822_cast_fp16, y = var_7823_cast_fp16)[name = string("op_7825_cast_fp16")];
            int32 var_7827 = const()[name = string("op_7827"), val = int32(-1)];
            bool q_43_interleave_0 = const()[name = string("q_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_43_cast_fp16 = concat(axis = var_7827, interleave = q_43_interleave_0, values = (var_7821_cast_fp16, var_7825_cast_fp16))[name = string("q_43_cast_fp16")];
            tensor<int32, [4]> first_87_begin_0 = const()[name = string("first_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_87_end_0 = const()[name = string("first_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_87_end_mask_0 = const()[name = string("first_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_87_cast_fp16 = slice_by_index(begin = first_87_begin_0, end = first_87_end_0, end_mask = first_87_end_mask_0, x = x_261_cast_fp16)[name = string("first_87_cast_fp16")];
            tensor<int32, [4]> second_87_begin_0 = const()[name = string("second_87_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_87_end_0 = const()[name = string("second_87_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_87_end_mask_0 = const()[name = string("second_87_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_87_cast_fp16 = slice_by_index(begin = second_87_begin_0, end = second_87_end_0, end_mask = second_87_end_mask_0, x = x_261_cast_fp16)[name = string("second_87_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7839_cast_fp16 = mul(x = first_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7839_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7840_cast_fp16 = mul(x = second_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7840_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7842_cast_fp16 = sub(x = var_7839_cast_fp16, y = var_7840_cast_fp16)[name = string("op_7842_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7843_cast_fp16 = mul(x = second_87_cast_fp16, y = cos_1_cast_fp16)[name = string("op_7843_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7844_cast_fp16 = mul(x = first_87_cast_fp16, y = sin_1_cast_fp16)[name = string("op_7844_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_7846_cast_fp16 = add(x = var_7843_cast_fp16, y = var_7844_cast_fp16)[name = string("op_7846_cast_fp16")];
            int32 var_7848 = const()[name = string("op_7848"), val = int32(-1)];
            bool k_43_interleave_0 = const()[name = string("k_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_43_cast_fp16 = concat(axis = var_7848, interleave = k_43_interleave_0, values = (var_7842_cast_fp16, var_7846_cast_fp16))[name = string("k_43_cast_fp16")];
            tensor<int32, [4]> var_7859_begin_0 = const()[name = string("op_7859_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7859_end_0 = const()[name = string("op_7859_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7859_end_mask_0 = const()[name = string("op_7859_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7859_stride_0 = const()[name = string("op_7859_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7859_cast_fp16 = slice_by_index(begin = var_7859_begin_0, end = var_7859_end_0, end_mask = var_7859_end_mask_0, stride = var_7859_stride_0, x = q_43_cast_fp16)[name = string("op_7859_cast_fp16")];
            tensor<int32, [4]> var_7869_begin_0 = const()[name = string("op_7869_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_7869_end_0 = const()[name = string("op_7869_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_7869_end_mask_0 = const()[name = string("op_7869_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_7869_stride_0 = const()[name = string("op_7869_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_7869_cast_fp16 = slice_by_index(begin = var_7869_begin_0, end = var_7869_end_0, end_mask = var_7869_end_mask_0, stride = var_7869_stride_0, x = q_43_cast_fp16)[name = string("op_7869_cast_fp16")];
            int32 var_7871 = const()[name = string("op_7871"), val = int32(2)];
            bool stacked_43_interleave_0 = const()[name = string("stacked_43_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_43_cast_fp16 = concat(axis = var_7871, interleave = stacked_43_interleave_0, values = (var_7859_cast_fp16, var_7869_cast_fp16))[name = string("stacked_43_cast_fp16")];
            tensor<int32, [4]> var_7877_begin_0 = const()[name = string("op_7877_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7877_end_0 = const()[name = string("op_7877_end_0"), val = tensor<int32, [4]>([22, 8, 528, 128])];
            tensor<bool, [4]> var_7877_end_mask_0 = const()[name = string("op_7877_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7877_cast_fp16 = slice_by_index(begin = var_7877_begin_0, end = var_7877_end_0, end_mask = var_7877_end_mask_0, x = read_state_0)[name = string("op_7877_cast_fp16")];
            tensor<int32, [4]> var_7887_begin_0 = const()[name = string("op_7887_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7887_end_0 = const()[name = string("op_7887_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7887_end_mask_0 = const()[name = string("op_7887_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7887_cast_fp16 = slice_by_index(begin = var_7887_begin_0, end = var_7887_end_0, end_mask = var_7887_end_mask_0, x = var_7877_cast_fp16)[name = string("op_7887_cast_fp16")];
            bool var_7896_transpose_x_1 = const()[name = string("op_7896_transpose_x_1"), val = bool(false)];
            bool var_7896_transpose_y_1 = const()[name = string("op_7896_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_7896_cast_fp16 = matmul(transpose_x = var_7896_transpose_x_1, transpose_y = var_7896_transpose_y_1, x = var_7887_cast_fp16, y = stacked_43_cast_fp16)[name = string("op_7896_cast_fp16")];
            tensor<int32, [4]> var_7899_perm_0 = const()[name = string("op_7899_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_7900_to_fp16 = const()[name = string("op_7900_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_7899_cast_fp16 = transpose(perm = var_7899_perm_0, x = var_7896_cast_fp16)[name = string("transpose_32")];
            tensor<fp16, [1, 8, 32, 512]> var_7901_cast_fp16 = mul(x = var_7899_cast_fp16, y = var_7900_to_fp16)[name = string("op_7901_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_43_cast_fp16 = add(x = var_7901_cast_fp16, y = cacheMask)[name = string("cache_weights_43_cast_fp16")];
            bool var_7907_transpose_x_1 = const()[name = string("op_7907_transpose_x_1"), val = bool(false)];
            bool var_7907_transpose_y_1 = const()[name = string("op_7907_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_7907_cast_fp16 = matmul(transpose_x = var_7907_transpose_x_1, transpose_y = var_7907_transpose_y_1, x = stacked_43_cast_fp16, y = k_43_cast_fp16)[name = string("op_7907_cast_fp16")];
            fp16 var_7908_to_fp16 = const()[name = string("op_7908_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_7909_cast_fp16 = mul(x = var_7907_cast_fp16, y = var_7908_to_fp16)[name = string("op_7909_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_43_cast_fp16 = add(x = var_7909_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_43_cast_fp16")];
            int32 var_7913 = const()[name = string("op_7913"), val = int32(-1)];
            bool var_7914_interleave_0 = const()[name = string("op_7914_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_7914_cast_fp16 = concat(axis = var_7913, interleave = var_7914_interleave_0, values = (cache_weights_43_cast_fp16, call_weights_43_cast_fp16))[name = string("op_7914_cast_fp16")];
            int32 var_7915 = const()[name = string("op_7915"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_43_cast_fp16 = softmax(axis = var_7915, x = var_7914_cast_fp16)[name = string("shares_43_cast_fp16")];
            tensor<int32, [4]> var_7922_begin_0 = const()[name = string("op_7922_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7922_end_0 = const()[name = string("op_7922_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_7922_end_mask_0 = const()[name = string("op_7922_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_7922_cast_fp16 = slice_by_index(begin = var_7922_begin_0, end = var_7922_end_0, end_mask = var_7922_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7922_cast_fp16")];
            tensor<int32, [4]> var_7927_begin_0 = const()[name = string("op_7927_begin_0"), val = tensor<int32, [4]>([21, 0, 0, 0])];
            tensor<int32, [4]> var_7927_end_0 = const()[name = string("op_7927_end_0"), val = tensor<int32, [4]>([22, 8, 528, 128])];
            tensor<bool, [4]> var_7927_end_mask_0 = const()[name = string("op_7927_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_7927_cast_fp16 = slice_by_index(begin = var_7927_begin_0, end = var_7927_end_0, end_mask = var_7927_end_mask_0, x = read_state_1)[name = string("op_7927_cast_fp16")];
            tensor<int32, [4]> var_7937_begin_0 = const()[name = string("op_7937_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7937_end_0 = const()[name = string("op_7937_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_7937_end_mask_0 = const()[name = string("op_7937_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_7937_cast_fp16 = slice_by_index(begin = var_7937_begin_0, end = var_7937_end_0, end_mask = var_7937_end_mask_0, x = var_7927_cast_fp16)[name = string("op_7937_cast_fp16")];
            bool var_7943_transpose_x_0 = const()[name = string("op_7943_transpose_x_0"), val = bool(false)];
            bool var_7943_transpose_y_0 = const()[name = string("op_7943_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_7943_cast_fp16 = matmul(transpose_x = var_7943_transpose_x_0, transpose_y = var_7943_transpose_y_0, x = var_7922_cast_fp16, y = var_7937_cast_fp16)[name = string("op_7943_cast_fp16")];
            tensor<int32, [4]> var_7948_begin_0 = const()[name = string("op_7948_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_7948_end_0 = const()[name = string("op_7948_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_7948_end_mask_0 = const()[name = string("op_7948_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_7948_cast_fp16 = slice_by_index(begin = var_7948_begin_0, end = var_7948_end_0, end_mask = var_7948_end_mask_0, x = shares_43_cast_fp16)[name = string("op_7948_cast_fp16")];
            bool var_7949_transpose_x_0 = const()[name = string("op_7949_transpose_x_0"), val = bool(false)];
            bool var_7949_transpose_y_0 = const()[name = string("op_7949_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_43_cast_fp16 = transpose(perm = v_43_perm_0, x = var_7804_cast_fp16)[name = string("transpose_31")];
            tensor<fp16, [1, 8, 32, 128]> var_7949_cast_fp16 = matmul(transpose_x = var_7949_transpose_x_0, transpose_y = var_7949_transpose_y_0, x = var_7948_cast_fp16, y = v_43_cast_fp16)[name = string("op_7949_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_85_cast_fp16 = add(x = var_7943_cast_fp16, y = var_7949_cast_fp16)[name = string("mixed_85_cast_fp16")];
            tensor<int32, [4]> var_7966_begin_0 = const()[name = string("op_7966_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_7966_end_0 = const()[name = string("op_7966_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_7966_end_mask_0 = const()[name = string("op_7966_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7966_cast_fp16 = slice_by_index(begin = var_7966_begin_0, end = var_7966_end_0, end_mask = var_7966_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7966_cast_fp16")];
            tensor<int32, [4]> var_7981_begin_0 = const()[name = string("op_7981_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_7981_end_0 = const()[name = string("op_7981_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_7981_end_mask_0 = const()[name = string("op_7981_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_7981_cast_fp16 = slice_by_index(begin = var_7981_begin_0, end = var_7981_end_0, end_mask = var_7981_end_mask_0, x = mixed_85_cast_fp16)[name = string("op_7981_cast_fp16")];
            int32 var_7983 = const()[name = string("op_7983"), val = int32(1)];
            bool mixed_87_interleave_0 = const()[name = string("mixed_87_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_87_cast_fp16 = concat(axis = var_7983, interleave = mixed_87_interleave_0, values = (var_7966_cast_fp16, var_7981_cast_fp16))[name = string("mixed_87_cast_fp16")];
            tensor<int32, [4]> var_7987_perm_0 = const()[name = string("op_7987_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_7991 = const()[name = string("op_7991"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_7987_cast_fp16 = transpose(perm = var_7987_perm_0, x = mixed_87_cast_fp16)[name = string("transpose_30")];
            tensor<fp16, [1, 16, 2048]> input_213_cast_fp16 = reshape(shape = var_7991, x = var_7987_cast_fp16)[name = string("input_213_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_21_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(646392768))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648489984))))[name = string("model_layers_21_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_150_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_self_attn_o_proj_weight_to_fp16_quantized, x = input_213_cast_fp16)[name = string("linear_150_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_263_cast_fp16 = add(x = x_253_cast_fp16, y = linear_150_cast_fp16)[name = string("x_263_cast_fp16")];
            fp16 var_8002_promoted_to_fp16 = const()[name = string("op_8002_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8004_cast_fp16 = pow(x = x_263_cast_fp16, y = var_8002_promoted_to_fp16)[name = string("op_8004_cast_fp16")];
            tensor<int32, [1]> var_8006_axes_0 = const()[name = string("op_8006_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8006_keep_dims_0 = const()[name = string("op_8006_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8006_cast_fp16 = reduce_mean(axes = var_8006_axes_0, keep_dims = var_8006_keep_dims_0, x = var_8004_cast_fp16)[name = string("op_8006_cast_fp16")];
            fp16 var_8007_to_fp16 = const()[name = string("op_8007_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8008_cast_fp16 = add(x = var_8006_cast_fp16, y = var_8007_to_fp16)[name = string("op_8008_cast_fp16")];
            fp32 var_8009_epsilon_0 = const()[name = string("op_8009_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8009_cast_fp16 = rsqrt(epsilon = var_8009_epsilon_0, x = var_8008_cast_fp16)[name = string("op_8009_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8010_cast_fp16 = mul(x = x_263_cast_fp16, y = var_8009_cast_fp16)[name = string("op_8010_cast_fp16")];
            tensor<fp16, [1024]> model_layers_21_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_21_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648492096)))];
            tensor<fp16, [1, 16, 1024]> input_215_cast_fp16 = mul(x = var_8010_cast_fp16, y = model_layers_21_post_attention_layernorm_weight_to_fp16)[name = string("input_215_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(648494208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651640000))))[name = string("model_layers_21_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_151_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_gate_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_151_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8015_cast_fp16 = silu(x = linear_151_cast_fp16)[name = string("op_8015_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_21_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(651646208))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654792000))))[name = string("model_layers_21_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_152_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_21_mlp_up_proj_weight_to_fp16_quantized, x = input_215_cast_fp16)[name = string("linear_152_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_219_cast_fp16 = mul(x = var_8015_cast_fp16, y = linear_152_cast_fp16)[name = string("input_219_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_21_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(654798208))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657944000))))[name = string("model_layers_21_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_153_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_21_mlp_down_proj_weight_to_fp16_quantized, x = input_219_cast_fp16)[name = string("linear_153_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_265_cast_fp16 = add(x = x_263_cast_fp16, y = linear_153_cast_fp16)[name = string("x_265_cast_fp16")];
            fp16 var_8029_promoted_to_fp16 = const()[name = string("op_8029_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8031_cast_fp16 = pow(x = x_265_cast_fp16, y = var_8029_promoted_to_fp16)[name = string("op_8031_cast_fp16")];
            tensor<int32, [1]> var_8033_axes_0 = const()[name = string("op_8033_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8033_keep_dims_0 = const()[name = string("op_8033_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8033_cast_fp16 = reduce_mean(axes = var_8033_axes_0, keep_dims = var_8033_keep_dims_0, x = var_8031_cast_fp16)[name = string("op_8033_cast_fp16")];
            fp16 var_8034_to_fp16 = const()[name = string("op_8034_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8035_cast_fp16 = add(x = var_8033_cast_fp16, y = var_8034_to_fp16)[name = string("op_8035_cast_fp16")];
            fp32 var_8036_epsilon_0 = const()[name = string("op_8036_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8036_cast_fp16 = rsqrt(epsilon = var_8036_epsilon_0, x = var_8035_cast_fp16)[name = string("op_8036_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8037_cast_fp16 = mul(x = x_265_cast_fp16, y = var_8036_cast_fp16)[name = string("op_8037_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657946112)))];
            tensor<fp16, [1, 16, 1024]> input_221_cast_fp16 = mul(x = var_8037_cast_fp16, y = model_layers_22_input_layernorm_weight_to_fp16)[name = string("input_221_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_22_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(657948224))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660045440))))[name = string("model_layers_22_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_154_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_22_self_attn_q_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_154_cast_fp16")];
            tensor<int32, [4]> var_8046 = const()[name = string("op_8046"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8047_cast_fp16 = reshape(shape = var_8046, x = linear_154_cast_fp16)[name = string("op_8047_cast_fp16")];
            tensor<int32, [4]> x_267_perm_0 = const()[name = string("x_267_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8055_promoted_to_fp16 = const()[name = string("op_8055_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_267_cast_fp16 = transpose(perm = x_267_perm_0, x = var_8047_cast_fp16)[name = string("transpose_29")];
            tensor<fp16, [1, 16, 16, 128]> var_8057_cast_fp16 = pow(x = x_267_cast_fp16, y = var_8055_promoted_to_fp16)[name = string("op_8057_cast_fp16")];
            tensor<int32, [1]> var_8059_axes_0 = const()[name = string("op_8059_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8059_keep_dims_0 = const()[name = string("op_8059_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8059_cast_fp16 = reduce_mean(axes = var_8059_axes_0, keep_dims = var_8059_keep_dims_0, x = var_8057_cast_fp16)[name = string("op_8059_cast_fp16")];
            fp16 var_8060_to_fp16 = const()[name = string("op_8060_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8061_cast_fp16 = add(x = var_8059_cast_fp16, y = var_8060_to_fp16)[name = string("op_8061_cast_fp16")];
            fp32 var_8062_epsilon_0 = const()[name = string("op_8062_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8062_cast_fp16 = rsqrt(epsilon = var_8062_epsilon_0, x = var_8061_cast_fp16)[name = string("op_8062_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8063_cast_fp16 = mul(x = x_267_cast_fp16, y = var_8062_cast_fp16)[name = string("op_8063_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049600)))];
            tensor<fp16, [1, 16, 16, 128]> x_271_cast_fp16 = mul(x = var_8063_cast_fp16, y = model_layers_22_self_attn_q_norm_weight_to_fp16)[name = string("x_271_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(660049920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661098560))))[name = string("model_layers_22_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_155_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_k_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_155_cast_fp16")];
            tensor<int32, [4]> var_8072 = const()[name = string("op_8072"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8073_cast_fp16 = reshape(shape = var_8072, x = linear_155_cast_fp16)[name = string("op_8073_cast_fp16")];
            tensor<int32, [4]> x_269_perm_0 = const()[name = string("x_269_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8081_promoted_to_fp16 = const()[name = string("op_8081_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_269_cast_fp16 = transpose(perm = x_269_perm_0, x = var_8073_cast_fp16)[name = string("transpose_28")];
            tensor<fp16, [1, 8, 16, 128]> var_8083_cast_fp16 = pow(x = x_269_cast_fp16, y = var_8081_promoted_to_fp16)[name = string("op_8083_cast_fp16")];
            tensor<int32, [1]> var_8085_axes_0 = const()[name = string("op_8085_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8085_keep_dims_0 = const()[name = string("op_8085_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8085_cast_fp16 = reduce_mean(axes = var_8085_axes_0, keep_dims = var_8085_keep_dims_0, x = var_8083_cast_fp16)[name = string("op_8085_cast_fp16")];
            fp16 var_8086_to_fp16 = const()[name = string("op_8086_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8087_cast_fp16 = add(x = var_8085_cast_fp16, y = var_8086_to_fp16)[name = string("op_8087_cast_fp16")];
            fp32 var_8088_epsilon_0 = const()[name = string("op_8088_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8088_cast_fp16 = rsqrt(epsilon = var_8088_epsilon_0, x = var_8087_cast_fp16)[name = string("op_8088_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8089_cast_fp16 = mul(x = x_269_cast_fp16, y = var_8088_cast_fp16)[name = string("op_8089_cast_fp16")];
            tensor<fp16, [128]> model_layers_22_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_22_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100672)))];
            tensor<fp16, [1, 8, 16, 128]> x_273_cast_fp16 = mul(x = var_8089_cast_fp16, y = model_layers_22_self_attn_k_norm_weight_to_fp16)[name = string("x_273_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_22_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(661100992))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662149632))))[name = string("model_layers_22_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_156_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_v_proj_weight_to_fp16_quantized, x = input_221_cast_fp16)[name = string("linear_156_cast_fp16")];
            tensor<int32, [4]> var_8098 = const()[name = string("op_8098"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8099_cast_fp16 = reshape(shape = var_8098, x = linear_156_cast_fp16)[name = string("op_8099_cast_fp16")];
            tensor<int32, [4]> v_45_perm_0 = const()[name = string("v_45_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_89_begin_0 = const()[name = string("first_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_89_end_0 = const()[name = string("first_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_89_end_mask_0 = const()[name = string("first_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_89_cast_fp16 = slice_by_index(begin = first_89_begin_0, end = first_89_end_0, end_mask = first_89_end_mask_0, x = x_271_cast_fp16)[name = string("first_89_cast_fp16")];
            tensor<int32, [4]> second_89_begin_0 = const()[name = string("second_89_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_89_end_0 = const()[name = string("second_89_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_89_end_mask_0 = const()[name = string("second_89_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_89_cast_fp16 = slice_by_index(begin = second_89_begin_0, end = second_89_end_0, end_mask = second_89_end_mask_0, x = x_271_cast_fp16)[name = string("second_89_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8113_cast_fp16 = mul(x = first_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8113_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8114_cast_fp16 = mul(x = second_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8114_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8116_cast_fp16 = sub(x = var_8113_cast_fp16, y = var_8114_cast_fp16)[name = string("op_8116_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8117_cast_fp16 = mul(x = second_89_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8117_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8118_cast_fp16 = mul(x = first_89_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8118_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8120_cast_fp16 = add(x = var_8117_cast_fp16, y = var_8118_cast_fp16)[name = string("op_8120_cast_fp16")];
            int32 var_8122 = const()[name = string("op_8122"), val = int32(-1)];
            bool q_45_interleave_0 = const()[name = string("q_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_45_cast_fp16 = concat(axis = var_8122, interleave = q_45_interleave_0, values = (var_8116_cast_fp16, var_8120_cast_fp16))[name = string("q_45_cast_fp16")];
            tensor<int32, [4]> first_91_begin_0 = const()[name = string("first_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_91_end_0 = const()[name = string("first_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_91_end_mask_0 = const()[name = string("first_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_91_cast_fp16 = slice_by_index(begin = first_91_begin_0, end = first_91_end_0, end_mask = first_91_end_mask_0, x = x_273_cast_fp16)[name = string("first_91_cast_fp16")];
            tensor<int32, [4]> second_91_begin_0 = const()[name = string("second_91_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_91_end_0 = const()[name = string("second_91_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_91_end_mask_0 = const()[name = string("second_91_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_91_cast_fp16 = slice_by_index(begin = second_91_begin_0, end = second_91_end_0, end_mask = second_91_end_mask_0, x = x_273_cast_fp16)[name = string("second_91_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8134_cast_fp16 = mul(x = first_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8134_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8135_cast_fp16 = mul(x = second_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8135_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8137_cast_fp16 = sub(x = var_8134_cast_fp16, y = var_8135_cast_fp16)[name = string("op_8137_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8138_cast_fp16 = mul(x = second_91_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8138_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8139_cast_fp16 = mul(x = first_91_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8139_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8141_cast_fp16 = add(x = var_8138_cast_fp16, y = var_8139_cast_fp16)[name = string("op_8141_cast_fp16")];
            int32 var_8143 = const()[name = string("op_8143"), val = int32(-1)];
            bool k_45_interleave_0 = const()[name = string("k_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_45_cast_fp16 = concat(axis = var_8143, interleave = k_45_interleave_0, values = (var_8137_cast_fp16, var_8141_cast_fp16))[name = string("k_45_cast_fp16")];
            tensor<int32, [4]> var_8154_begin_0 = const()[name = string("op_8154_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8154_end_0 = const()[name = string("op_8154_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8154_end_mask_0 = const()[name = string("op_8154_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8154_stride_0 = const()[name = string("op_8154_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8154_cast_fp16 = slice_by_index(begin = var_8154_begin_0, end = var_8154_end_0, end_mask = var_8154_end_mask_0, stride = var_8154_stride_0, x = q_45_cast_fp16)[name = string("op_8154_cast_fp16")];
            tensor<int32, [4]> var_8164_begin_0 = const()[name = string("op_8164_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8164_end_0 = const()[name = string("op_8164_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8164_end_mask_0 = const()[name = string("op_8164_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8164_stride_0 = const()[name = string("op_8164_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8164_cast_fp16 = slice_by_index(begin = var_8164_begin_0, end = var_8164_end_0, end_mask = var_8164_end_mask_0, stride = var_8164_stride_0, x = q_45_cast_fp16)[name = string("op_8164_cast_fp16")];
            int32 var_8166 = const()[name = string("op_8166"), val = int32(2)];
            bool stacked_45_interleave_0 = const()[name = string("stacked_45_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_45_cast_fp16 = concat(axis = var_8166, interleave = stacked_45_interleave_0, values = (var_8154_cast_fp16, var_8164_cast_fp16))[name = string("stacked_45_cast_fp16")];
            tensor<int32, [4]> var_8172_begin_0 = const()[name = string("op_8172_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8172_end_0 = const()[name = string("op_8172_end_0"), val = tensor<int32, [4]>([23, 8, 528, 128])];
            tensor<bool, [4]> var_8172_end_mask_0 = const()[name = string("op_8172_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8172_cast_fp16 = slice_by_index(begin = var_8172_begin_0, end = var_8172_end_0, end_mask = var_8172_end_mask_0, x = read_state_0)[name = string("op_8172_cast_fp16")];
            tensor<int32, [4]> var_8182_begin_0 = const()[name = string("op_8182_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8182_end_0 = const()[name = string("op_8182_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8182_end_mask_0 = const()[name = string("op_8182_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8182_cast_fp16 = slice_by_index(begin = var_8182_begin_0, end = var_8182_end_0, end_mask = var_8182_end_mask_0, x = var_8172_cast_fp16)[name = string("op_8182_cast_fp16")];
            bool var_8191_transpose_x_1 = const()[name = string("op_8191_transpose_x_1"), val = bool(false)];
            bool var_8191_transpose_y_1 = const()[name = string("op_8191_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8191_cast_fp16 = matmul(transpose_x = var_8191_transpose_x_1, transpose_y = var_8191_transpose_y_1, x = var_8182_cast_fp16, y = stacked_45_cast_fp16)[name = string("op_8191_cast_fp16")];
            tensor<int32, [4]> var_8194_perm_0 = const()[name = string("op_8194_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8195_to_fp16 = const()[name = string("op_8195_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8194_cast_fp16 = transpose(perm = var_8194_perm_0, x = var_8191_cast_fp16)[name = string("transpose_27")];
            tensor<fp16, [1, 8, 32, 512]> var_8196_cast_fp16 = mul(x = var_8194_cast_fp16, y = var_8195_to_fp16)[name = string("op_8196_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_45_cast_fp16 = add(x = var_8196_cast_fp16, y = cacheMask)[name = string("cache_weights_45_cast_fp16")];
            bool var_8202_transpose_x_1 = const()[name = string("op_8202_transpose_x_1"), val = bool(false)];
            bool var_8202_transpose_y_1 = const()[name = string("op_8202_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8202_cast_fp16 = matmul(transpose_x = var_8202_transpose_x_1, transpose_y = var_8202_transpose_y_1, x = stacked_45_cast_fp16, y = k_45_cast_fp16)[name = string("op_8202_cast_fp16")];
            fp16 var_8203_to_fp16 = const()[name = string("op_8203_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8204_cast_fp16 = mul(x = var_8202_cast_fp16, y = var_8203_to_fp16)[name = string("op_8204_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_45_cast_fp16 = add(x = var_8204_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_45_cast_fp16")];
            int32 var_8208 = const()[name = string("op_8208"), val = int32(-1)];
            bool var_8209_interleave_0 = const()[name = string("op_8209_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8209_cast_fp16 = concat(axis = var_8208, interleave = var_8209_interleave_0, values = (cache_weights_45_cast_fp16, call_weights_45_cast_fp16))[name = string("op_8209_cast_fp16")];
            int32 var_8210 = const()[name = string("op_8210"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_45_cast_fp16 = softmax(axis = var_8210, x = var_8209_cast_fp16)[name = string("shares_45_cast_fp16")];
            tensor<int32, [4]> var_8217_begin_0 = const()[name = string("op_8217_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8217_end_0 = const()[name = string("op_8217_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8217_end_mask_0 = const()[name = string("op_8217_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8217_cast_fp16 = slice_by_index(begin = var_8217_begin_0, end = var_8217_end_0, end_mask = var_8217_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8217_cast_fp16")];
            tensor<int32, [4]> var_8222_begin_0 = const()[name = string("op_8222_begin_0"), val = tensor<int32, [4]>([22, 0, 0, 0])];
            tensor<int32, [4]> var_8222_end_0 = const()[name = string("op_8222_end_0"), val = tensor<int32, [4]>([23, 8, 528, 128])];
            tensor<bool, [4]> var_8222_end_mask_0 = const()[name = string("op_8222_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8222_cast_fp16 = slice_by_index(begin = var_8222_begin_0, end = var_8222_end_0, end_mask = var_8222_end_mask_0, x = read_state_1)[name = string("op_8222_cast_fp16")];
            tensor<int32, [4]> var_8232_begin_0 = const()[name = string("op_8232_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8232_end_0 = const()[name = string("op_8232_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8232_end_mask_0 = const()[name = string("op_8232_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8232_cast_fp16 = slice_by_index(begin = var_8232_begin_0, end = var_8232_end_0, end_mask = var_8232_end_mask_0, x = var_8222_cast_fp16)[name = string("op_8232_cast_fp16")];
            bool var_8238_transpose_x_0 = const()[name = string("op_8238_transpose_x_0"), val = bool(false)];
            bool var_8238_transpose_y_0 = const()[name = string("op_8238_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8238_cast_fp16 = matmul(transpose_x = var_8238_transpose_x_0, transpose_y = var_8238_transpose_y_0, x = var_8217_cast_fp16, y = var_8232_cast_fp16)[name = string("op_8238_cast_fp16")];
            tensor<int32, [4]> var_8243_begin_0 = const()[name = string("op_8243_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8243_end_0 = const()[name = string("op_8243_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8243_end_mask_0 = const()[name = string("op_8243_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8243_cast_fp16 = slice_by_index(begin = var_8243_begin_0, end = var_8243_end_0, end_mask = var_8243_end_mask_0, x = shares_45_cast_fp16)[name = string("op_8243_cast_fp16")];
            bool var_8244_transpose_x_0 = const()[name = string("op_8244_transpose_x_0"), val = bool(false)];
            bool var_8244_transpose_y_0 = const()[name = string("op_8244_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_45_cast_fp16 = transpose(perm = v_45_perm_0, x = var_8099_cast_fp16)[name = string("transpose_26")];
            tensor<fp16, [1, 8, 32, 128]> var_8244_cast_fp16 = matmul(transpose_x = var_8244_transpose_x_0, transpose_y = var_8244_transpose_y_0, x = var_8243_cast_fp16, y = v_45_cast_fp16)[name = string("op_8244_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_89_cast_fp16 = add(x = var_8238_cast_fp16, y = var_8244_cast_fp16)[name = string("mixed_89_cast_fp16")];
            tensor<int32, [4]> var_8261_begin_0 = const()[name = string("op_8261_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8261_end_0 = const()[name = string("op_8261_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8261_end_mask_0 = const()[name = string("op_8261_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8261_cast_fp16 = slice_by_index(begin = var_8261_begin_0, end = var_8261_end_0, end_mask = var_8261_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8261_cast_fp16")];
            tensor<int32, [4]> var_8276_begin_0 = const()[name = string("op_8276_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8276_end_0 = const()[name = string("op_8276_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8276_end_mask_0 = const()[name = string("op_8276_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8276_cast_fp16 = slice_by_index(begin = var_8276_begin_0, end = var_8276_end_0, end_mask = var_8276_end_mask_0, x = mixed_89_cast_fp16)[name = string("op_8276_cast_fp16")];
            int32 var_8278 = const()[name = string("op_8278"), val = int32(1)];
            bool mixed_91_interleave_0 = const()[name = string("mixed_91_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_91_cast_fp16 = concat(axis = var_8278, interleave = mixed_91_interleave_0, values = (var_8261_cast_fp16, var_8276_cast_fp16))[name = string("mixed_91_cast_fp16")];
            tensor<int32, [4]> var_8282_perm_0 = const()[name = string("op_8282_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8286 = const()[name = string("op_8286"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8282_cast_fp16 = transpose(perm = var_8282_perm_0, x = mixed_91_cast_fp16)[name = string("transpose_25")];
            tensor<fp16, [1, 16, 2048]> input_223_cast_fp16 = reshape(shape = var_8286, x = var_8282_cast_fp16)[name = string("input_223_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_22_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(662151744))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664248960))))[name = string("model_layers_22_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_157_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_self_attn_o_proj_weight_to_fp16_quantized, x = input_223_cast_fp16)[name = string("linear_157_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_275_cast_fp16 = add(x = x_265_cast_fp16, y = linear_157_cast_fp16)[name = string("x_275_cast_fp16")];
            fp16 var_8297_promoted_to_fp16 = const()[name = string("op_8297_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8299_cast_fp16 = pow(x = x_275_cast_fp16, y = var_8297_promoted_to_fp16)[name = string("op_8299_cast_fp16")];
            tensor<int32, [1]> var_8301_axes_0 = const()[name = string("op_8301_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8301_keep_dims_0 = const()[name = string("op_8301_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8301_cast_fp16 = reduce_mean(axes = var_8301_axes_0, keep_dims = var_8301_keep_dims_0, x = var_8299_cast_fp16)[name = string("op_8301_cast_fp16")];
            fp16 var_8302_to_fp16 = const()[name = string("op_8302_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8303_cast_fp16 = add(x = var_8301_cast_fp16, y = var_8302_to_fp16)[name = string("op_8303_cast_fp16")];
            fp32 var_8304_epsilon_0 = const()[name = string("op_8304_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8304_cast_fp16 = rsqrt(epsilon = var_8304_epsilon_0, x = var_8303_cast_fp16)[name = string("op_8304_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8305_cast_fp16 = mul(x = x_275_cast_fp16, y = var_8304_cast_fp16)[name = string("op_8305_cast_fp16")];
            tensor<fp16, [1024]> model_layers_22_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_22_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664251072)))];
            tensor<fp16, [1, 16, 1024]> input_225_cast_fp16 = mul(x = var_8305_cast_fp16, y = model_layers_22_post_attention_layernorm_weight_to_fp16)[name = string("input_225_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(664253184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667398976))))[name = string("model_layers_22_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_158_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_gate_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_158_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8310_cast_fp16 = silu(x = linear_158_cast_fp16)[name = string("op_8310_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_22_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(667405184))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670550976))))[name = string("model_layers_22_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_159_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_22_mlp_up_proj_weight_to_fp16_quantized, x = input_225_cast_fp16)[name = string("linear_159_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_229_cast_fp16 = mul(x = var_8310_cast_fp16, y = linear_159_cast_fp16)[name = string("input_229_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_22_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(670557184))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673702976))))[name = string("model_layers_22_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_160_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_22_mlp_down_proj_weight_to_fp16_quantized, x = input_229_cast_fp16)[name = string("linear_160_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_277_cast_fp16 = add(x = x_275_cast_fp16, y = linear_160_cast_fp16)[name = string("x_277_cast_fp16")];
            fp16 var_8324_promoted_to_fp16 = const()[name = string("op_8324_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8326_cast_fp16 = pow(x = x_277_cast_fp16, y = var_8324_promoted_to_fp16)[name = string("op_8326_cast_fp16")];
            tensor<int32, [1]> var_8328_axes_0 = const()[name = string("op_8328_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8328_keep_dims_0 = const()[name = string("op_8328_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8328_cast_fp16 = reduce_mean(axes = var_8328_axes_0, keep_dims = var_8328_keep_dims_0, x = var_8326_cast_fp16)[name = string("op_8328_cast_fp16")];
            fp16 var_8329_to_fp16 = const()[name = string("op_8329_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8330_cast_fp16 = add(x = var_8328_cast_fp16, y = var_8329_to_fp16)[name = string("op_8330_cast_fp16")];
            fp32 var_8331_epsilon_0 = const()[name = string("op_8331_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8331_cast_fp16 = rsqrt(epsilon = var_8331_epsilon_0, x = var_8330_cast_fp16)[name = string("op_8331_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8332_cast_fp16 = mul(x = x_277_cast_fp16, y = var_8331_cast_fp16)[name = string("op_8332_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673705088)))];
            tensor<fp16, [1, 16, 1024]> input_231_cast_fp16 = mul(x = var_8332_cast_fp16, y = model_layers_23_input_layernorm_weight_to_fp16)[name = string("input_231_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_23_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(673707200))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675804416))))[name = string("model_layers_23_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_161_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_23_self_attn_q_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_161_cast_fp16")];
            tensor<int32, [4]> var_8341 = const()[name = string("op_8341"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8342_cast_fp16 = reshape(shape = var_8341, x = linear_161_cast_fp16)[name = string("op_8342_cast_fp16")];
            tensor<int32, [4]> x_279_perm_0 = const()[name = string("x_279_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8350_promoted_to_fp16 = const()[name = string("op_8350_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_279_cast_fp16 = transpose(perm = x_279_perm_0, x = var_8342_cast_fp16)[name = string("transpose_24")];
            tensor<fp16, [1, 16, 16, 128]> var_8352_cast_fp16 = pow(x = x_279_cast_fp16, y = var_8350_promoted_to_fp16)[name = string("op_8352_cast_fp16")];
            tensor<int32, [1]> var_8354_axes_0 = const()[name = string("op_8354_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8354_keep_dims_0 = const()[name = string("op_8354_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8354_cast_fp16 = reduce_mean(axes = var_8354_axes_0, keep_dims = var_8354_keep_dims_0, x = var_8352_cast_fp16)[name = string("op_8354_cast_fp16")];
            fp16 var_8355_to_fp16 = const()[name = string("op_8355_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8356_cast_fp16 = add(x = var_8354_cast_fp16, y = var_8355_to_fp16)[name = string("op_8356_cast_fp16")];
            fp32 var_8357_epsilon_0 = const()[name = string("op_8357_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8357_cast_fp16 = rsqrt(epsilon = var_8357_epsilon_0, x = var_8356_cast_fp16)[name = string("op_8357_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8358_cast_fp16 = mul(x = x_279_cast_fp16, y = var_8357_cast_fp16)[name = string("op_8358_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808576)))];
            tensor<fp16, [1, 16, 16, 128]> x_283_cast_fp16 = mul(x = var_8358_cast_fp16, y = model_layers_23_self_attn_q_norm_weight_to_fp16)[name = string("x_283_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(675808896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676857536))))[name = string("model_layers_23_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_162_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_k_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_162_cast_fp16")];
            tensor<int32, [4]> var_8367 = const()[name = string("op_8367"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8368_cast_fp16 = reshape(shape = var_8367, x = linear_162_cast_fp16)[name = string("op_8368_cast_fp16")];
            tensor<int32, [4]> x_281_perm_0 = const()[name = string("x_281_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8376_promoted_to_fp16 = const()[name = string("op_8376_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_281_cast_fp16 = transpose(perm = x_281_perm_0, x = var_8368_cast_fp16)[name = string("transpose_23")];
            tensor<fp16, [1, 8, 16, 128]> var_8378_cast_fp16 = pow(x = x_281_cast_fp16, y = var_8376_promoted_to_fp16)[name = string("op_8378_cast_fp16")];
            tensor<int32, [1]> var_8380_axes_0 = const()[name = string("op_8380_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8380_keep_dims_0 = const()[name = string("op_8380_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8380_cast_fp16 = reduce_mean(axes = var_8380_axes_0, keep_dims = var_8380_keep_dims_0, x = var_8378_cast_fp16)[name = string("op_8380_cast_fp16")];
            fp16 var_8381_to_fp16 = const()[name = string("op_8381_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8382_cast_fp16 = add(x = var_8380_cast_fp16, y = var_8381_to_fp16)[name = string("op_8382_cast_fp16")];
            fp32 var_8383_epsilon_0 = const()[name = string("op_8383_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8383_cast_fp16 = rsqrt(epsilon = var_8383_epsilon_0, x = var_8382_cast_fp16)[name = string("op_8383_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8384_cast_fp16 = mul(x = x_281_cast_fp16, y = var_8383_cast_fp16)[name = string("op_8384_cast_fp16")];
            tensor<fp16, [128]> model_layers_23_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_23_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859648)))];
            tensor<fp16, [1, 8, 16, 128]> x_285_cast_fp16 = mul(x = var_8384_cast_fp16, y = model_layers_23_self_attn_k_norm_weight_to_fp16)[name = string("x_285_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_23_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(676859968))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677908608))))[name = string("model_layers_23_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_163_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_v_proj_weight_to_fp16_quantized, x = input_231_cast_fp16)[name = string("linear_163_cast_fp16")];
            tensor<int32, [4]> var_8393 = const()[name = string("op_8393"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8394_cast_fp16 = reshape(shape = var_8393, x = linear_163_cast_fp16)[name = string("op_8394_cast_fp16")];
            tensor<int32, [4]> v_47_perm_0 = const()[name = string("v_47_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_93_begin_0 = const()[name = string("first_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_93_end_0 = const()[name = string("first_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_93_end_mask_0 = const()[name = string("first_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_93_cast_fp16 = slice_by_index(begin = first_93_begin_0, end = first_93_end_0, end_mask = first_93_end_mask_0, x = x_283_cast_fp16)[name = string("first_93_cast_fp16")];
            tensor<int32, [4]> second_93_begin_0 = const()[name = string("second_93_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_93_end_0 = const()[name = string("second_93_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_93_end_mask_0 = const()[name = string("second_93_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_93_cast_fp16 = slice_by_index(begin = second_93_begin_0, end = second_93_end_0, end_mask = second_93_end_mask_0, x = x_283_cast_fp16)[name = string("second_93_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8408_cast_fp16 = mul(x = first_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8408_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8409_cast_fp16 = mul(x = second_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8409_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8411_cast_fp16 = sub(x = var_8408_cast_fp16, y = var_8409_cast_fp16)[name = string("op_8411_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8412_cast_fp16 = mul(x = second_93_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8412_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8413_cast_fp16 = mul(x = first_93_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8413_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8415_cast_fp16 = add(x = var_8412_cast_fp16, y = var_8413_cast_fp16)[name = string("op_8415_cast_fp16")];
            int32 var_8417 = const()[name = string("op_8417"), val = int32(-1)];
            bool q_47_interleave_0 = const()[name = string("q_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_47_cast_fp16 = concat(axis = var_8417, interleave = q_47_interleave_0, values = (var_8411_cast_fp16, var_8415_cast_fp16))[name = string("q_47_cast_fp16")];
            tensor<int32, [4]> first_95_begin_0 = const()[name = string("first_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_95_end_0 = const()[name = string("first_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_95_end_mask_0 = const()[name = string("first_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_95_cast_fp16 = slice_by_index(begin = first_95_begin_0, end = first_95_end_0, end_mask = first_95_end_mask_0, x = x_285_cast_fp16)[name = string("first_95_cast_fp16")];
            tensor<int32, [4]> second_95_begin_0 = const()[name = string("second_95_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_95_end_0 = const()[name = string("second_95_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_95_end_mask_0 = const()[name = string("second_95_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_95_cast_fp16 = slice_by_index(begin = second_95_begin_0, end = second_95_end_0, end_mask = second_95_end_mask_0, x = x_285_cast_fp16)[name = string("second_95_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8429_cast_fp16 = mul(x = first_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8429_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8430_cast_fp16 = mul(x = second_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8430_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8432_cast_fp16 = sub(x = var_8429_cast_fp16, y = var_8430_cast_fp16)[name = string("op_8432_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8433_cast_fp16 = mul(x = second_95_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8433_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8434_cast_fp16 = mul(x = first_95_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8434_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8436_cast_fp16 = add(x = var_8433_cast_fp16, y = var_8434_cast_fp16)[name = string("op_8436_cast_fp16")];
            int32 var_8438 = const()[name = string("op_8438"), val = int32(-1)];
            bool k_47_interleave_0 = const()[name = string("k_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_47_cast_fp16 = concat(axis = var_8438, interleave = k_47_interleave_0, values = (var_8432_cast_fp16, var_8436_cast_fp16))[name = string("k_47_cast_fp16")];
            tensor<int32, [4]> var_8449_begin_0 = const()[name = string("op_8449_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8449_end_0 = const()[name = string("op_8449_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8449_end_mask_0 = const()[name = string("op_8449_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8449_stride_0 = const()[name = string("op_8449_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8449_cast_fp16 = slice_by_index(begin = var_8449_begin_0, end = var_8449_end_0, end_mask = var_8449_end_mask_0, stride = var_8449_stride_0, x = q_47_cast_fp16)[name = string("op_8449_cast_fp16")];
            tensor<int32, [4]> var_8459_begin_0 = const()[name = string("op_8459_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8459_end_0 = const()[name = string("op_8459_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8459_end_mask_0 = const()[name = string("op_8459_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8459_stride_0 = const()[name = string("op_8459_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8459_cast_fp16 = slice_by_index(begin = var_8459_begin_0, end = var_8459_end_0, end_mask = var_8459_end_mask_0, stride = var_8459_stride_0, x = q_47_cast_fp16)[name = string("op_8459_cast_fp16")];
            int32 var_8461 = const()[name = string("op_8461"), val = int32(2)];
            bool stacked_47_interleave_0 = const()[name = string("stacked_47_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_47_cast_fp16 = concat(axis = var_8461, interleave = stacked_47_interleave_0, values = (var_8449_cast_fp16, var_8459_cast_fp16))[name = string("stacked_47_cast_fp16")];
            tensor<int32, [4]> var_8467_begin_0 = const()[name = string("op_8467_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8467_end_0 = const()[name = string("op_8467_end_0"), val = tensor<int32, [4]>([24, 8, 528, 128])];
            tensor<bool, [4]> var_8467_end_mask_0 = const()[name = string("op_8467_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8467_cast_fp16 = slice_by_index(begin = var_8467_begin_0, end = var_8467_end_0, end_mask = var_8467_end_mask_0, x = read_state_0)[name = string("op_8467_cast_fp16")];
            tensor<int32, [4]> var_8477_begin_0 = const()[name = string("op_8477_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8477_end_0 = const()[name = string("op_8477_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8477_end_mask_0 = const()[name = string("op_8477_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8477_cast_fp16 = slice_by_index(begin = var_8477_begin_0, end = var_8477_end_0, end_mask = var_8477_end_mask_0, x = var_8467_cast_fp16)[name = string("op_8477_cast_fp16")];
            bool var_8486_transpose_x_1 = const()[name = string("op_8486_transpose_x_1"), val = bool(false)];
            bool var_8486_transpose_y_1 = const()[name = string("op_8486_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8486_cast_fp16 = matmul(transpose_x = var_8486_transpose_x_1, transpose_y = var_8486_transpose_y_1, x = var_8477_cast_fp16, y = stacked_47_cast_fp16)[name = string("op_8486_cast_fp16")];
            tensor<int32, [4]> var_8489_perm_0 = const()[name = string("op_8489_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8490_to_fp16 = const()[name = string("op_8490_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8489_cast_fp16 = transpose(perm = var_8489_perm_0, x = var_8486_cast_fp16)[name = string("transpose_22")];
            tensor<fp16, [1, 8, 32, 512]> var_8491_cast_fp16 = mul(x = var_8489_cast_fp16, y = var_8490_to_fp16)[name = string("op_8491_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_47_cast_fp16 = add(x = var_8491_cast_fp16, y = cacheMask)[name = string("cache_weights_47_cast_fp16")];
            bool var_8497_transpose_x_1 = const()[name = string("op_8497_transpose_x_1"), val = bool(false)];
            bool var_8497_transpose_y_1 = const()[name = string("op_8497_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8497_cast_fp16 = matmul(transpose_x = var_8497_transpose_x_1, transpose_y = var_8497_transpose_y_1, x = stacked_47_cast_fp16, y = k_47_cast_fp16)[name = string("op_8497_cast_fp16")];
            fp16 var_8498_to_fp16 = const()[name = string("op_8498_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8499_cast_fp16 = mul(x = var_8497_cast_fp16, y = var_8498_to_fp16)[name = string("op_8499_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_47_cast_fp16 = add(x = var_8499_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_47_cast_fp16")];
            int32 var_8503 = const()[name = string("op_8503"), val = int32(-1)];
            bool var_8504_interleave_0 = const()[name = string("op_8504_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8504_cast_fp16 = concat(axis = var_8503, interleave = var_8504_interleave_0, values = (cache_weights_47_cast_fp16, call_weights_47_cast_fp16))[name = string("op_8504_cast_fp16")];
            int32 var_8505 = const()[name = string("op_8505"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_47_cast_fp16 = softmax(axis = var_8505, x = var_8504_cast_fp16)[name = string("shares_47_cast_fp16")];
            tensor<int32, [4]> var_8512_begin_0 = const()[name = string("op_8512_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8512_end_0 = const()[name = string("op_8512_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8512_end_mask_0 = const()[name = string("op_8512_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8512_cast_fp16 = slice_by_index(begin = var_8512_begin_0, end = var_8512_end_0, end_mask = var_8512_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8512_cast_fp16")];
            tensor<int32, [4]> var_8517_begin_0 = const()[name = string("op_8517_begin_0"), val = tensor<int32, [4]>([23, 0, 0, 0])];
            tensor<int32, [4]> var_8517_end_0 = const()[name = string("op_8517_end_0"), val = tensor<int32, [4]>([24, 8, 528, 128])];
            tensor<bool, [4]> var_8517_end_mask_0 = const()[name = string("op_8517_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8517_cast_fp16 = slice_by_index(begin = var_8517_begin_0, end = var_8517_end_0, end_mask = var_8517_end_mask_0, x = read_state_1)[name = string("op_8517_cast_fp16")];
            tensor<int32, [4]> var_8527_begin_0 = const()[name = string("op_8527_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8527_end_0 = const()[name = string("op_8527_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8527_end_mask_0 = const()[name = string("op_8527_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8527_cast_fp16 = slice_by_index(begin = var_8527_begin_0, end = var_8527_end_0, end_mask = var_8527_end_mask_0, x = var_8517_cast_fp16)[name = string("op_8527_cast_fp16")];
            bool var_8533_transpose_x_0 = const()[name = string("op_8533_transpose_x_0"), val = bool(false)];
            bool var_8533_transpose_y_0 = const()[name = string("op_8533_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8533_cast_fp16 = matmul(transpose_x = var_8533_transpose_x_0, transpose_y = var_8533_transpose_y_0, x = var_8512_cast_fp16, y = var_8527_cast_fp16)[name = string("op_8533_cast_fp16")];
            tensor<int32, [4]> var_8538_begin_0 = const()[name = string("op_8538_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8538_end_0 = const()[name = string("op_8538_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8538_end_mask_0 = const()[name = string("op_8538_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8538_cast_fp16 = slice_by_index(begin = var_8538_begin_0, end = var_8538_end_0, end_mask = var_8538_end_mask_0, x = shares_47_cast_fp16)[name = string("op_8538_cast_fp16")];
            bool var_8539_transpose_x_0 = const()[name = string("op_8539_transpose_x_0"), val = bool(false)];
            bool var_8539_transpose_y_0 = const()[name = string("op_8539_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_47_cast_fp16 = transpose(perm = v_47_perm_0, x = var_8394_cast_fp16)[name = string("transpose_21")];
            tensor<fp16, [1, 8, 32, 128]> var_8539_cast_fp16 = matmul(transpose_x = var_8539_transpose_x_0, transpose_y = var_8539_transpose_y_0, x = var_8538_cast_fp16, y = v_47_cast_fp16)[name = string("op_8539_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_93_cast_fp16 = add(x = var_8533_cast_fp16, y = var_8539_cast_fp16)[name = string("mixed_93_cast_fp16")];
            tensor<int32, [4]> var_8556_begin_0 = const()[name = string("op_8556_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8556_end_0 = const()[name = string("op_8556_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8556_end_mask_0 = const()[name = string("op_8556_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8556_cast_fp16 = slice_by_index(begin = var_8556_begin_0, end = var_8556_end_0, end_mask = var_8556_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8556_cast_fp16")];
            tensor<int32, [4]> var_8571_begin_0 = const()[name = string("op_8571_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8571_end_0 = const()[name = string("op_8571_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8571_end_mask_0 = const()[name = string("op_8571_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8571_cast_fp16 = slice_by_index(begin = var_8571_begin_0, end = var_8571_end_0, end_mask = var_8571_end_mask_0, x = mixed_93_cast_fp16)[name = string("op_8571_cast_fp16")];
            int32 var_8573 = const()[name = string("op_8573"), val = int32(1)];
            bool mixed_95_interleave_0 = const()[name = string("mixed_95_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_95_cast_fp16 = concat(axis = var_8573, interleave = mixed_95_interleave_0, values = (var_8556_cast_fp16, var_8571_cast_fp16))[name = string("mixed_95_cast_fp16")];
            tensor<int32, [4]> var_8577_perm_0 = const()[name = string("op_8577_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8581 = const()[name = string("op_8581"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8577_cast_fp16 = transpose(perm = var_8577_perm_0, x = mixed_95_cast_fp16)[name = string("transpose_20")];
            tensor<fp16, [1, 16, 2048]> input_233_cast_fp16 = reshape(shape = var_8581, x = var_8577_cast_fp16)[name = string("input_233_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_23_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(677910720))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680007936))))[name = string("model_layers_23_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_164_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_self_attn_o_proj_weight_to_fp16_quantized, x = input_233_cast_fp16)[name = string("linear_164_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_287_cast_fp16 = add(x = x_277_cast_fp16, y = linear_164_cast_fp16)[name = string("x_287_cast_fp16")];
            fp16 var_8592_promoted_to_fp16 = const()[name = string("op_8592_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8594_cast_fp16 = pow(x = x_287_cast_fp16, y = var_8592_promoted_to_fp16)[name = string("op_8594_cast_fp16")];
            tensor<int32, [1]> var_8596_axes_0 = const()[name = string("op_8596_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8596_keep_dims_0 = const()[name = string("op_8596_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8596_cast_fp16 = reduce_mean(axes = var_8596_axes_0, keep_dims = var_8596_keep_dims_0, x = var_8594_cast_fp16)[name = string("op_8596_cast_fp16")];
            fp16 var_8597_to_fp16 = const()[name = string("op_8597_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8598_cast_fp16 = add(x = var_8596_cast_fp16, y = var_8597_to_fp16)[name = string("op_8598_cast_fp16")];
            fp32 var_8599_epsilon_0 = const()[name = string("op_8599_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8599_cast_fp16 = rsqrt(epsilon = var_8599_epsilon_0, x = var_8598_cast_fp16)[name = string("op_8599_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8600_cast_fp16 = mul(x = x_287_cast_fp16, y = var_8599_cast_fp16)[name = string("op_8600_cast_fp16")];
            tensor<fp16, [1024]> model_layers_23_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_23_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680010048)))];
            tensor<fp16, [1, 16, 1024]> input_235_cast_fp16 = mul(x = var_8600_cast_fp16, y = model_layers_23_post_attention_layernorm_weight_to_fp16)[name = string("input_235_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(680012160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683157952))))[name = string("model_layers_23_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_165_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_gate_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_165_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8605_cast_fp16 = silu(x = linear_165_cast_fp16)[name = string("op_8605_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_23_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(683164160))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686309952))))[name = string("model_layers_23_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_166_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_23_mlp_up_proj_weight_to_fp16_quantized, x = input_235_cast_fp16)[name = string("linear_166_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_239_cast_fp16 = mul(x = var_8605_cast_fp16, y = linear_166_cast_fp16)[name = string("input_239_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_23_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(686316160))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689461952))))[name = string("model_layers_23_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_167_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_23_mlp_down_proj_weight_to_fp16_quantized, x = input_239_cast_fp16)[name = string("linear_167_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_289_cast_fp16 = add(x = x_287_cast_fp16, y = linear_167_cast_fp16)[name = string("x_289_cast_fp16")];
            fp16 var_8619_promoted_to_fp16 = const()[name = string("op_8619_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8621_cast_fp16 = pow(x = x_289_cast_fp16, y = var_8619_promoted_to_fp16)[name = string("op_8621_cast_fp16")];
            tensor<int32, [1]> var_8623_axes_0 = const()[name = string("op_8623_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8623_keep_dims_0 = const()[name = string("op_8623_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8623_cast_fp16 = reduce_mean(axes = var_8623_axes_0, keep_dims = var_8623_keep_dims_0, x = var_8621_cast_fp16)[name = string("op_8623_cast_fp16")];
            fp16 var_8624_to_fp16 = const()[name = string("op_8624_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8625_cast_fp16 = add(x = var_8623_cast_fp16, y = var_8624_to_fp16)[name = string("op_8625_cast_fp16")];
            fp32 var_8626_epsilon_0 = const()[name = string("op_8626_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8626_cast_fp16 = rsqrt(epsilon = var_8626_epsilon_0, x = var_8625_cast_fp16)[name = string("op_8626_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8627_cast_fp16 = mul(x = x_289_cast_fp16, y = var_8626_cast_fp16)[name = string("op_8627_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689464064)))];
            tensor<fp16, [1, 16, 1024]> input_241_cast_fp16 = mul(x = var_8627_cast_fp16, y = model_layers_24_input_layernorm_weight_to_fp16)[name = string("input_241_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_24_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(689466176))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691563392))))[name = string("model_layers_24_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_168_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_24_self_attn_q_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_168_cast_fp16")];
            tensor<int32, [4]> var_8636 = const()[name = string("op_8636"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8637_cast_fp16 = reshape(shape = var_8636, x = linear_168_cast_fp16)[name = string("op_8637_cast_fp16")];
            tensor<int32, [4]> x_291_perm_0 = const()[name = string("x_291_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8645_promoted_to_fp16 = const()[name = string("op_8645_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_291_cast_fp16 = transpose(perm = x_291_perm_0, x = var_8637_cast_fp16)[name = string("transpose_19")];
            tensor<fp16, [1, 16, 16, 128]> var_8647_cast_fp16 = pow(x = x_291_cast_fp16, y = var_8645_promoted_to_fp16)[name = string("op_8647_cast_fp16")];
            tensor<int32, [1]> var_8649_axes_0 = const()[name = string("op_8649_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8649_keep_dims_0 = const()[name = string("op_8649_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8649_cast_fp16 = reduce_mean(axes = var_8649_axes_0, keep_dims = var_8649_keep_dims_0, x = var_8647_cast_fp16)[name = string("op_8649_cast_fp16")];
            fp16 var_8650_to_fp16 = const()[name = string("op_8650_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8651_cast_fp16 = add(x = var_8649_cast_fp16, y = var_8650_to_fp16)[name = string("op_8651_cast_fp16")];
            fp32 var_8652_epsilon_0 = const()[name = string("op_8652_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8652_cast_fp16 = rsqrt(epsilon = var_8652_epsilon_0, x = var_8651_cast_fp16)[name = string("op_8652_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8653_cast_fp16 = mul(x = x_291_cast_fp16, y = var_8652_cast_fp16)[name = string("op_8653_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567552)))];
            tensor<fp16, [1, 16, 16, 128]> x_295_cast_fp16 = mul(x = var_8653_cast_fp16, y = model_layers_24_self_attn_q_norm_weight_to_fp16)[name = string("x_295_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(691567872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692616512))))[name = string("model_layers_24_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_169_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_k_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_169_cast_fp16")];
            tensor<int32, [4]> var_8662 = const()[name = string("op_8662"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8663_cast_fp16 = reshape(shape = var_8662, x = linear_169_cast_fp16)[name = string("op_8663_cast_fp16")];
            tensor<int32, [4]> x_293_perm_0 = const()[name = string("x_293_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8671_promoted_to_fp16 = const()[name = string("op_8671_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_293_cast_fp16 = transpose(perm = x_293_perm_0, x = var_8663_cast_fp16)[name = string("transpose_18")];
            tensor<fp16, [1, 8, 16, 128]> var_8673_cast_fp16 = pow(x = x_293_cast_fp16, y = var_8671_promoted_to_fp16)[name = string("op_8673_cast_fp16")];
            tensor<int32, [1]> var_8675_axes_0 = const()[name = string("op_8675_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8675_keep_dims_0 = const()[name = string("op_8675_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8675_cast_fp16 = reduce_mean(axes = var_8675_axes_0, keep_dims = var_8675_keep_dims_0, x = var_8673_cast_fp16)[name = string("op_8675_cast_fp16")];
            fp16 var_8676_to_fp16 = const()[name = string("op_8676_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8677_cast_fp16 = add(x = var_8675_cast_fp16, y = var_8676_to_fp16)[name = string("op_8677_cast_fp16")];
            fp32 var_8678_epsilon_0 = const()[name = string("op_8678_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8678_cast_fp16 = rsqrt(epsilon = var_8678_epsilon_0, x = var_8677_cast_fp16)[name = string("op_8678_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8679_cast_fp16 = mul(x = x_293_cast_fp16, y = var_8678_cast_fp16)[name = string("op_8679_cast_fp16")];
            tensor<fp16, [128]> model_layers_24_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_24_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618624)))];
            tensor<fp16, [1, 8, 16, 128]> x_297_cast_fp16 = mul(x = var_8679_cast_fp16, y = model_layers_24_self_attn_k_norm_weight_to_fp16)[name = string("x_297_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_24_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(692618944))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693667584))))[name = string("model_layers_24_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_170_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_v_proj_weight_to_fp16_quantized, x = input_241_cast_fp16)[name = string("linear_170_cast_fp16")];
            tensor<int32, [4]> var_8688 = const()[name = string("op_8688"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8689_cast_fp16 = reshape(shape = var_8688, x = linear_170_cast_fp16)[name = string("op_8689_cast_fp16")];
            tensor<int32, [4]> v_49_perm_0 = const()[name = string("v_49_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_97_begin_0 = const()[name = string("first_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_97_end_0 = const()[name = string("first_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_97_end_mask_0 = const()[name = string("first_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_97_cast_fp16 = slice_by_index(begin = first_97_begin_0, end = first_97_end_0, end_mask = first_97_end_mask_0, x = x_295_cast_fp16)[name = string("first_97_cast_fp16")];
            tensor<int32, [4]> second_97_begin_0 = const()[name = string("second_97_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_97_end_0 = const()[name = string("second_97_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_97_end_mask_0 = const()[name = string("second_97_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_97_cast_fp16 = slice_by_index(begin = second_97_begin_0, end = second_97_end_0, end_mask = second_97_end_mask_0, x = x_295_cast_fp16)[name = string("second_97_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8703_cast_fp16 = mul(x = first_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8703_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8704_cast_fp16 = mul(x = second_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8704_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8706_cast_fp16 = sub(x = var_8703_cast_fp16, y = var_8704_cast_fp16)[name = string("op_8706_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8707_cast_fp16 = mul(x = second_97_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8707_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8708_cast_fp16 = mul(x = first_97_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8708_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8710_cast_fp16 = add(x = var_8707_cast_fp16, y = var_8708_cast_fp16)[name = string("op_8710_cast_fp16")];
            int32 var_8712 = const()[name = string("op_8712"), val = int32(-1)];
            bool q_49_interleave_0 = const()[name = string("q_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_49_cast_fp16 = concat(axis = var_8712, interleave = q_49_interleave_0, values = (var_8706_cast_fp16, var_8710_cast_fp16))[name = string("q_49_cast_fp16")];
            tensor<int32, [4]> first_99_begin_0 = const()[name = string("first_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_99_end_0 = const()[name = string("first_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_99_end_mask_0 = const()[name = string("first_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_99_cast_fp16 = slice_by_index(begin = first_99_begin_0, end = first_99_end_0, end_mask = first_99_end_mask_0, x = x_297_cast_fp16)[name = string("first_99_cast_fp16")];
            tensor<int32, [4]> second_99_begin_0 = const()[name = string("second_99_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_99_end_0 = const()[name = string("second_99_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_99_end_mask_0 = const()[name = string("second_99_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_99_cast_fp16 = slice_by_index(begin = second_99_begin_0, end = second_99_end_0, end_mask = second_99_end_mask_0, x = x_297_cast_fp16)[name = string("second_99_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8724_cast_fp16 = mul(x = first_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8724_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8725_cast_fp16 = mul(x = second_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8725_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8727_cast_fp16 = sub(x = var_8724_cast_fp16, y = var_8725_cast_fp16)[name = string("op_8727_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8728_cast_fp16 = mul(x = second_99_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8728_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8729_cast_fp16 = mul(x = first_99_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8729_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_8731_cast_fp16 = add(x = var_8728_cast_fp16, y = var_8729_cast_fp16)[name = string("op_8731_cast_fp16")];
            int32 var_8733 = const()[name = string("op_8733"), val = int32(-1)];
            bool k_49_interleave_0 = const()[name = string("k_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_49_cast_fp16 = concat(axis = var_8733, interleave = k_49_interleave_0, values = (var_8727_cast_fp16, var_8731_cast_fp16))[name = string("k_49_cast_fp16")];
            tensor<int32, [4]> var_8744_begin_0 = const()[name = string("op_8744_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8744_end_0 = const()[name = string("op_8744_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8744_end_mask_0 = const()[name = string("op_8744_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8744_stride_0 = const()[name = string("op_8744_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8744_cast_fp16 = slice_by_index(begin = var_8744_begin_0, end = var_8744_end_0, end_mask = var_8744_end_mask_0, stride = var_8744_stride_0, x = q_49_cast_fp16)[name = string("op_8744_cast_fp16")];
            tensor<int32, [4]> var_8754_begin_0 = const()[name = string("op_8754_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_8754_end_0 = const()[name = string("op_8754_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_8754_end_mask_0 = const()[name = string("op_8754_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_8754_stride_0 = const()[name = string("op_8754_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_8754_cast_fp16 = slice_by_index(begin = var_8754_begin_0, end = var_8754_end_0, end_mask = var_8754_end_mask_0, stride = var_8754_stride_0, x = q_49_cast_fp16)[name = string("op_8754_cast_fp16")];
            int32 var_8756 = const()[name = string("op_8756"), val = int32(2)];
            bool stacked_49_interleave_0 = const()[name = string("stacked_49_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_49_cast_fp16 = concat(axis = var_8756, interleave = stacked_49_interleave_0, values = (var_8744_cast_fp16, var_8754_cast_fp16))[name = string("stacked_49_cast_fp16")];
            tensor<int32, [4]> var_8762_begin_0 = const()[name = string("op_8762_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8762_end_0 = const()[name = string("op_8762_end_0"), val = tensor<int32, [4]>([25, 8, 528, 128])];
            tensor<bool, [4]> var_8762_end_mask_0 = const()[name = string("op_8762_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8762_cast_fp16 = slice_by_index(begin = var_8762_begin_0, end = var_8762_end_0, end_mask = var_8762_end_mask_0, x = read_state_0)[name = string("op_8762_cast_fp16")];
            tensor<int32, [4]> var_8772_begin_0 = const()[name = string("op_8772_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8772_end_0 = const()[name = string("op_8772_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8772_end_mask_0 = const()[name = string("op_8772_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8772_cast_fp16 = slice_by_index(begin = var_8772_begin_0, end = var_8772_end_0, end_mask = var_8772_end_mask_0, x = var_8762_cast_fp16)[name = string("op_8772_cast_fp16")];
            bool var_8781_transpose_x_1 = const()[name = string("op_8781_transpose_x_1"), val = bool(false)];
            bool var_8781_transpose_y_1 = const()[name = string("op_8781_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_8781_cast_fp16 = matmul(transpose_x = var_8781_transpose_x_1, transpose_y = var_8781_transpose_y_1, x = var_8772_cast_fp16, y = stacked_49_cast_fp16)[name = string("op_8781_cast_fp16")];
            tensor<int32, [4]> var_8784_perm_0 = const()[name = string("op_8784_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_8785_to_fp16 = const()[name = string("op_8785_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_8784_cast_fp16 = transpose(perm = var_8784_perm_0, x = var_8781_cast_fp16)[name = string("transpose_17")];
            tensor<fp16, [1, 8, 32, 512]> var_8786_cast_fp16 = mul(x = var_8784_cast_fp16, y = var_8785_to_fp16)[name = string("op_8786_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_49_cast_fp16 = add(x = var_8786_cast_fp16, y = cacheMask)[name = string("cache_weights_49_cast_fp16")];
            bool var_8792_transpose_x_1 = const()[name = string("op_8792_transpose_x_1"), val = bool(false)];
            bool var_8792_transpose_y_1 = const()[name = string("op_8792_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_8792_cast_fp16 = matmul(transpose_x = var_8792_transpose_x_1, transpose_y = var_8792_transpose_y_1, x = stacked_49_cast_fp16, y = k_49_cast_fp16)[name = string("op_8792_cast_fp16")];
            fp16 var_8793_to_fp16 = const()[name = string("op_8793_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_8794_cast_fp16 = mul(x = var_8792_cast_fp16, y = var_8793_to_fp16)[name = string("op_8794_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_49_cast_fp16 = add(x = var_8794_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_49_cast_fp16")];
            int32 var_8798 = const()[name = string("op_8798"), val = int32(-1)];
            bool var_8799_interleave_0 = const()[name = string("op_8799_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_8799_cast_fp16 = concat(axis = var_8798, interleave = var_8799_interleave_0, values = (cache_weights_49_cast_fp16, call_weights_49_cast_fp16))[name = string("op_8799_cast_fp16")];
            int32 var_8800 = const()[name = string("op_8800"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_49_cast_fp16 = softmax(axis = var_8800, x = var_8799_cast_fp16)[name = string("shares_49_cast_fp16")];
            tensor<int32, [4]> var_8807_begin_0 = const()[name = string("op_8807_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8807_end_0 = const()[name = string("op_8807_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_8807_end_mask_0 = const()[name = string("op_8807_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_8807_cast_fp16 = slice_by_index(begin = var_8807_begin_0, end = var_8807_end_0, end_mask = var_8807_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8807_cast_fp16")];
            tensor<int32, [4]> var_8812_begin_0 = const()[name = string("op_8812_begin_0"), val = tensor<int32, [4]>([24, 0, 0, 0])];
            tensor<int32, [4]> var_8812_end_0 = const()[name = string("op_8812_end_0"), val = tensor<int32, [4]>([25, 8, 528, 128])];
            tensor<bool, [4]> var_8812_end_mask_0 = const()[name = string("op_8812_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_8812_cast_fp16 = slice_by_index(begin = var_8812_begin_0, end = var_8812_end_0, end_mask = var_8812_end_mask_0, x = read_state_1)[name = string("op_8812_cast_fp16")];
            tensor<int32, [4]> var_8822_begin_0 = const()[name = string("op_8822_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8822_end_0 = const()[name = string("op_8822_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_8822_end_mask_0 = const()[name = string("op_8822_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_8822_cast_fp16 = slice_by_index(begin = var_8822_begin_0, end = var_8822_end_0, end_mask = var_8822_end_mask_0, x = var_8812_cast_fp16)[name = string("op_8822_cast_fp16")];
            bool var_8828_transpose_x_0 = const()[name = string("op_8828_transpose_x_0"), val = bool(false)];
            bool var_8828_transpose_y_0 = const()[name = string("op_8828_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_8828_cast_fp16 = matmul(transpose_x = var_8828_transpose_x_0, transpose_y = var_8828_transpose_y_0, x = var_8807_cast_fp16, y = var_8822_cast_fp16)[name = string("op_8828_cast_fp16")];
            tensor<int32, [4]> var_8833_begin_0 = const()[name = string("op_8833_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_8833_end_0 = const()[name = string("op_8833_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_8833_end_mask_0 = const()[name = string("op_8833_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_8833_cast_fp16 = slice_by_index(begin = var_8833_begin_0, end = var_8833_end_0, end_mask = var_8833_end_mask_0, x = shares_49_cast_fp16)[name = string("op_8833_cast_fp16")];
            bool var_8834_transpose_x_0 = const()[name = string("op_8834_transpose_x_0"), val = bool(false)];
            bool var_8834_transpose_y_0 = const()[name = string("op_8834_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_49_cast_fp16 = transpose(perm = v_49_perm_0, x = var_8689_cast_fp16)[name = string("transpose_16")];
            tensor<fp16, [1, 8, 32, 128]> var_8834_cast_fp16 = matmul(transpose_x = var_8834_transpose_x_0, transpose_y = var_8834_transpose_y_0, x = var_8833_cast_fp16, y = v_49_cast_fp16)[name = string("op_8834_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_97_cast_fp16 = add(x = var_8828_cast_fp16, y = var_8834_cast_fp16)[name = string("mixed_97_cast_fp16")];
            tensor<int32, [4]> var_8851_begin_0 = const()[name = string("op_8851_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_8851_end_0 = const()[name = string("op_8851_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_8851_end_mask_0 = const()[name = string("op_8851_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8851_cast_fp16 = slice_by_index(begin = var_8851_begin_0, end = var_8851_end_0, end_mask = var_8851_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8851_cast_fp16")];
            tensor<int32, [4]> var_8866_begin_0 = const()[name = string("op_8866_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_8866_end_0 = const()[name = string("op_8866_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_8866_end_mask_0 = const()[name = string("op_8866_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_8866_cast_fp16 = slice_by_index(begin = var_8866_begin_0, end = var_8866_end_0, end_mask = var_8866_end_mask_0, x = mixed_97_cast_fp16)[name = string("op_8866_cast_fp16")];
            int32 var_8868 = const()[name = string("op_8868"), val = int32(1)];
            bool mixed_99_interleave_0 = const()[name = string("mixed_99_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_99_cast_fp16 = concat(axis = var_8868, interleave = mixed_99_interleave_0, values = (var_8851_cast_fp16, var_8866_cast_fp16))[name = string("mixed_99_cast_fp16")];
            tensor<int32, [4]> var_8872_perm_0 = const()[name = string("op_8872_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_8876 = const()[name = string("op_8876"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_8872_cast_fp16 = transpose(perm = var_8872_perm_0, x = mixed_99_cast_fp16)[name = string("transpose_15")];
            tensor<fp16, [1, 16, 2048]> input_243_cast_fp16 = reshape(shape = var_8876, x = var_8872_cast_fp16)[name = string("input_243_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_24_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(693669696))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695766912))))[name = string("model_layers_24_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_171_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_self_attn_o_proj_weight_to_fp16_quantized, x = input_243_cast_fp16)[name = string("linear_171_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_299_cast_fp16 = add(x = x_289_cast_fp16, y = linear_171_cast_fp16)[name = string("x_299_cast_fp16")];
            fp16 var_8887_promoted_to_fp16 = const()[name = string("op_8887_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8889_cast_fp16 = pow(x = x_299_cast_fp16, y = var_8887_promoted_to_fp16)[name = string("op_8889_cast_fp16")];
            tensor<int32, [1]> var_8891_axes_0 = const()[name = string("op_8891_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8891_keep_dims_0 = const()[name = string("op_8891_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8891_cast_fp16 = reduce_mean(axes = var_8891_axes_0, keep_dims = var_8891_keep_dims_0, x = var_8889_cast_fp16)[name = string("op_8891_cast_fp16")];
            fp16 var_8892_to_fp16 = const()[name = string("op_8892_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8893_cast_fp16 = add(x = var_8891_cast_fp16, y = var_8892_to_fp16)[name = string("op_8893_cast_fp16")];
            fp32 var_8894_epsilon_0 = const()[name = string("op_8894_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8894_cast_fp16 = rsqrt(epsilon = var_8894_epsilon_0, x = var_8893_cast_fp16)[name = string("op_8894_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8895_cast_fp16 = mul(x = x_299_cast_fp16, y = var_8894_cast_fp16)[name = string("op_8895_cast_fp16")];
            tensor<fp16, [1024]> model_layers_24_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_24_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695769024)))];
            tensor<fp16, [1, 16, 1024]> input_245_cast_fp16 = mul(x = var_8895_cast_fp16, y = model_layers_24_post_attention_layernorm_weight_to_fp16)[name = string("input_245_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(695771136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698916928))))[name = string("model_layers_24_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_172_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_gate_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_172_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_8900_cast_fp16 = silu(x = linear_172_cast_fp16)[name = string("op_8900_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_24_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(698923136))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702068928))))[name = string("model_layers_24_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_173_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_24_mlp_up_proj_weight_to_fp16_quantized, x = input_245_cast_fp16)[name = string("linear_173_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_249_cast_fp16 = mul(x = var_8900_cast_fp16, y = linear_173_cast_fp16)[name = string("input_249_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_24_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(702075136))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705220928))))[name = string("model_layers_24_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_174_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_24_mlp_down_proj_weight_to_fp16_quantized, x = input_249_cast_fp16)[name = string("linear_174_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_301_cast_fp16 = add(x = x_299_cast_fp16, y = linear_174_cast_fp16)[name = string("x_301_cast_fp16")];
            fp16 var_8914_promoted_to_fp16 = const()[name = string("op_8914_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_8916_cast_fp16 = pow(x = x_301_cast_fp16, y = var_8914_promoted_to_fp16)[name = string("op_8916_cast_fp16")];
            tensor<int32, [1]> var_8918_axes_0 = const()[name = string("op_8918_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8918_keep_dims_0 = const()[name = string("op_8918_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_8918_cast_fp16 = reduce_mean(axes = var_8918_axes_0, keep_dims = var_8918_keep_dims_0, x = var_8916_cast_fp16)[name = string("op_8918_cast_fp16")];
            fp16 var_8919_to_fp16 = const()[name = string("op_8919_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_8920_cast_fp16 = add(x = var_8918_cast_fp16, y = var_8919_to_fp16)[name = string("op_8920_cast_fp16")];
            fp32 var_8921_epsilon_0 = const()[name = string("op_8921_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_8921_cast_fp16 = rsqrt(epsilon = var_8921_epsilon_0, x = var_8920_cast_fp16)[name = string("op_8921_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_8922_cast_fp16 = mul(x = x_301_cast_fp16, y = var_8921_cast_fp16)[name = string("op_8922_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705223040)))];
            tensor<fp16, [1, 16, 1024]> input_251_cast_fp16 = mul(x = var_8922_cast_fp16, y = model_layers_25_input_layernorm_weight_to_fp16)[name = string("input_251_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_25_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(705225152))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707322368))))[name = string("model_layers_25_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_175_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_25_self_attn_q_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_175_cast_fp16")];
            tensor<int32, [4]> var_8931 = const()[name = string("op_8931"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_8932_cast_fp16 = reshape(shape = var_8931, x = linear_175_cast_fp16)[name = string("op_8932_cast_fp16")];
            tensor<int32, [4]> x_303_perm_0 = const()[name = string("x_303_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8940_promoted_to_fp16 = const()[name = string("op_8940_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_303_cast_fp16 = transpose(perm = x_303_perm_0, x = var_8932_cast_fp16)[name = string("transpose_14")];
            tensor<fp16, [1, 16, 16, 128]> var_8942_cast_fp16 = pow(x = x_303_cast_fp16, y = var_8940_promoted_to_fp16)[name = string("op_8942_cast_fp16")];
            tensor<int32, [1]> var_8944_axes_0 = const()[name = string("op_8944_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8944_keep_dims_0 = const()[name = string("op_8944_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_8944_cast_fp16 = reduce_mean(axes = var_8944_axes_0, keep_dims = var_8944_keep_dims_0, x = var_8942_cast_fp16)[name = string("op_8944_cast_fp16")];
            fp16 var_8945_to_fp16 = const()[name = string("op_8945_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_8946_cast_fp16 = add(x = var_8944_cast_fp16, y = var_8945_to_fp16)[name = string("op_8946_cast_fp16")];
            fp32 var_8947_epsilon_0 = const()[name = string("op_8947_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_8947_cast_fp16 = rsqrt(epsilon = var_8947_epsilon_0, x = var_8946_cast_fp16)[name = string("op_8947_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_8948_cast_fp16 = mul(x = x_303_cast_fp16, y = var_8947_cast_fp16)[name = string("op_8948_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326528)))];
            tensor<fp16, [1, 16, 16, 128]> x_307_cast_fp16 = mul(x = var_8948_cast_fp16, y = model_layers_25_self_attn_q_norm_weight_to_fp16)[name = string("x_307_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(707326848))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708375488))))[name = string("model_layers_25_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_176_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_k_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_176_cast_fp16")];
            tensor<int32, [4]> var_8957 = const()[name = string("op_8957"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8958_cast_fp16 = reshape(shape = var_8957, x = linear_176_cast_fp16)[name = string("op_8958_cast_fp16")];
            tensor<int32, [4]> x_305_perm_0 = const()[name = string("x_305_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_8966_promoted_to_fp16 = const()[name = string("op_8966_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_305_cast_fp16 = transpose(perm = x_305_perm_0, x = var_8958_cast_fp16)[name = string("transpose_13")];
            tensor<fp16, [1, 8, 16, 128]> var_8968_cast_fp16 = pow(x = x_305_cast_fp16, y = var_8966_promoted_to_fp16)[name = string("op_8968_cast_fp16")];
            tensor<int32, [1]> var_8970_axes_0 = const()[name = string("op_8970_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_8970_keep_dims_0 = const()[name = string("op_8970_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_8970_cast_fp16 = reduce_mean(axes = var_8970_axes_0, keep_dims = var_8970_keep_dims_0, x = var_8968_cast_fp16)[name = string("op_8970_cast_fp16")];
            fp16 var_8971_to_fp16 = const()[name = string("op_8971_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_8972_cast_fp16 = add(x = var_8970_cast_fp16, y = var_8971_to_fp16)[name = string("op_8972_cast_fp16")];
            fp32 var_8973_epsilon_0 = const()[name = string("op_8973_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_8973_cast_fp16 = rsqrt(epsilon = var_8973_epsilon_0, x = var_8972_cast_fp16)[name = string("op_8973_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_8974_cast_fp16 = mul(x = x_305_cast_fp16, y = var_8973_cast_fp16)[name = string("op_8974_cast_fp16")];
            tensor<fp16, [128]> model_layers_25_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_25_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377600)))];
            tensor<fp16, [1, 8, 16, 128]> x_309_cast_fp16 = mul(x = var_8974_cast_fp16, y = model_layers_25_self_attn_k_norm_weight_to_fp16)[name = string("x_309_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_25_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(708377920))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709426560))))[name = string("model_layers_25_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_177_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_v_proj_weight_to_fp16_quantized, x = input_251_cast_fp16)[name = string("linear_177_cast_fp16")];
            tensor<int32, [4]> var_8983 = const()[name = string("op_8983"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_8984_cast_fp16 = reshape(shape = var_8983, x = linear_177_cast_fp16)[name = string("op_8984_cast_fp16")];
            tensor<int32, [4]> v_51_perm_0 = const()[name = string("v_51_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_101_begin_0 = const()[name = string("first_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_101_end_0 = const()[name = string("first_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_101_end_mask_0 = const()[name = string("first_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_101_cast_fp16 = slice_by_index(begin = first_101_begin_0, end = first_101_end_0, end_mask = first_101_end_mask_0, x = x_307_cast_fp16)[name = string("first_101_cast_fp16")];
            tensor<int32, [4]> second_101_begin_0 = const()[name = string("second_101_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_101_end_0 = const()[name = string("second_101_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_101_end_mask_0 = const()[name = string("second_101_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_101_cast_fp16 = slice_by_index(begin = second_101_begin_0, end = second_101_end_0, end_mask = second_101_end_mask_0, x = x_307_cast_fp16)[name = string("second_101_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8998_cast_fp16 = mul(x = first_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_8998_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_8999_cast_fp16 = mul(x = second_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_8999_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9001_cast_fp16 = sub(x = var_8998_cast_fp16, y = var_8999_cast_fp16)[name = string("op_9001_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9002_cast_fp16 = mul(x = second_101_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9002_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9003_cast_fp16 = mul(x = first_101_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9003_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9005_cast_fp16 = add(x = var_9002_cast_fp16, y = var_9003_cast_fp16)[name = string("op_9005_cast_fp16")];
            int32 var_9007 = const()[name = string("op_9007"), val = int32(-1)];
            bool q_51_interleave_0 = const()[name = string("q_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_51_cast_fp16 = concat(axis = var_9007, interleave = q_51_interleave_0, values = (var_9001_cast_fp16, var_9005_cast_fp16))[name = string("q_51_cast_fp16")];
            tensor<int32, [4]> first_103_begin_0 = const()[name = string("first_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_103_end_0 = const()[name = string("first_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_103_end_mask_0 = const()[name = string("first_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_103_cast_fp16 = slice_by_index(begin = first_103_begin_0, end = first_103_end_0, end_mask = first_103_end_mask_0, x = x_309_cast_fp16)[name = string("first_103_cast_fp16")];
            tensor<int32, [4]> second_103_begin_0 = const()[name = string("second_103_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_103_end_0 = const()[name = string("second_103_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_103_end_mask_0 = const()[name = string("second_103_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_103_cast_fp16 = slice_by_index(begin = second_103_begin_0, end = second_103_end_0, end_mask = second_103_end_mask_0, x = x_309_cast_fp16)[name = string("second_103_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9019_cast_fp16 = mul(x = first_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9019_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9020_cast_fp16 = mul(x = second_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9020_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9022_cast_fp16 = sub(x = var_9019_cast_fp16, y = var_9020_cast_fp16)[name = string("op_9022_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9023_cast_fp16 = mul(x = second_103_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9023_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9024_cast_fp16 = mul(x = first_103_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9024_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9026_cast_fp16 = add(x = var_9023_cast_fp16, y = var_9024_cast_fp16)[name = string("op_9026_cast_fp16")];
            int32 var_9028 = const()[name = string("op_9028"), val = int32(-1)];
            bool k_51_interleave_0 = const()[name = string("k_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_51_cast_fp16 = concat(axis = var_9028, interleave = k_51_interleave_0, values = (var_9022_cast_fp16, var_9026_cast_fp16))[name = string("k_51_cast_fp16")];
            tensor<int32, [4]> var_9039_begin_0 = const()[name = string("op_9039_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9039_end_0 = const()[name = string("op_9039_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9039_end_mask_0 = const()[name = string("op_9039_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9039_stride_0 = const()[name = string("op_9039_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9039_cast_fp16 = slice_by_index(begin = var_9039_begin_0, end = var_9039_end_0, end_mask = var_9039_end_mask_0, stride = var_9039_stride_0, x = q_51_cast_fp16)[name = string("op_9039_cast_fp16")];
            tensor<int32, [4]> var_9049_begin_0 = const()[name = string("op_9049_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9049_end_0 = const()[name = string("op_9049_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9049_end_mask_0 = const()[name = string("op_9049_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9049_stride_0 = const()[name = string("op_9049_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9049_cast_fp16 = slice_by_index(begin = var_9049_begin_0, end = var_9049_end_0, end_mask = var_9049_end_mask_0, stride = var_9049_stride_0, x = q_51_cast_fp16)[name = string("op_9049_cast_fp16")];
            int32 var_9051 = const()[name = string("op_9051"), val = int32(2)];
            bool stacked_51_interleave_0 = const()[name = string("stacked_51_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_51_cast_fp16 = concat(axis = var_9051, interleave = stacked_51_interleave_0, values = (var_9039_cast_fp16, var_9049_cast_fp16))[name = string("stacked_51_cast_fp16")];
            tensor<int32, [4]> var_9057_begin_0 = const()[name = string("op_9057_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9057_end_0 = const()[name = string("op_9057_end_0"), val = tensor<int32, [4]>([26, 8, 528, 128])];
            tensor<bool, [4]> var_9057_end_mask_0 = const()[name = string("op_9057_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9057_cast_fp16 = slice_by_index(begin = var_9057_begin_0, end = var_9057_end_0, end_mask = var_9057_end_mask_0, x = read_state_0)[name = string("op_9057_cast_fp16")];
            tensor<int32, [4]> var_9067_begin_0 = const()[name = string("op_9067_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9067_end_0 = const()[name = string("op_9067_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9067_end_mask_0 = const()[name = string("op_9067_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9067_cast_fp16 = slice_by_index(begin = var_9067_begin_0, end = var_9067_end_0, end_mask = var_9067_end_mask_0, x = var_9057_cast_fp16)[name = string("op_9067_cast_fp16")];
            bool var_9076_transpose_x_1 = const()[name = string("op_9076_transpose_x_1"), val = bool(false)];
            bool var_9076_transpose_y_1 = const()[name = string("op_9076_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9076_cast_fp16 = matmul(transpose_x = var_9076_transpose_x_1, transpose_y = var_9076_transpose_y_1, x = var_9067_cast_fp16, y = stacked_51_cast_fp16)[name = string("op_9076_cast_fp16")];
            tensor<int32, [4]> var_9079_perm_0 = const()[name = string("op_9079_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9080_to_fp16 = const()[name = string("op_9080_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9079_cast_fp16 = transpose(perm = var_9079_perm_0, x = var_9076_cast_fp16)[name = string("transpose_12")];
            tensor<fp16, [1, 8, 32, 512]> var_9081_cast_fp16 = mul(x = var_9079_cast_fp16, y = var_9080_to_fp16)[name = string("op_9081_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_51_cast_fp16 = add(x = var_9081_cast_fp16, y = cacheMask)[name = string("cache_weights_51_cast_fp16")];
            bool var_9087_transpose_x_1 = const()[name = string("op_9087_transpose_x_1"), val = bool(false)];
            bool var_9087_transpose_y_1 = const()[name = string("op_9087_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9087_cast_fp16 = matmul(transpose_x = var_9087_transpose_x_1, transpose_y = var_9087_transpose_y_1, x = stacked_51_cast_fp16, y = k_51_cast_fp16)[name = string("op_9087_cast_fp16")];
            fp16 var_9088_to_fp16 = const()[name = string("op_9088_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9089_cast_fp16 = mul(x = var_9087_cast_fp16, y = var_9088_to_fp16)[name = string("op_9089_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_51_cast_fp16 = add(x = var_9089_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_51_cast_fp16")];
            int32 var_9093 = const()[name = string("op_9093"), val = int32(-1)];
            bool var_9094_interleave_0 = const()[name = string("op_9094_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9094_cast_fp16 = concat(axis = var_9093, interleave = var_9094_interleave_0, values = (cache_weights_51_cast_fp16, call_weights_51_cast_fp16))[name = string("op_9094_cast_fp16")];
            int32 var_9095 = const()[name = string("op_9095"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_51_cast_fp16 = softmax(axis = var_9095, x = var_9094_cast_fp16)[name = string("shares_51_cast_fp16")];
            tensor<int32, [4]> var_9102_begin_0 = const()[name = string("op_9102_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9102_end_0 = const()[name = string("op_9102_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9102_end_mask_0 = const()[name = string("op_9102_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9102_cast_fp16 = slice_by_index(begin = var_9102_begin_0, end = var_9102_end_0, end_mask = var_9102_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9102_cast_fp16")];
            tensor<int32, [4]> var_9107_begin_0 = const()[name = string("op_9107_begin_0"), val = tensor<int32, [4]>([25, 0, 0, 0])];
            tensor<int32, [4]> var_9107_end_0 = const()[name = string("op_9107_end_0"), val = tensor<int32, [4]>([26, 8, 528, 128])];
            tensor<bool, [4]> var_9107_end_mask_0 = const()[name = string("op_9107_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9107_cast_fp16 = slice_by_index(begin = var_9107_begin_0, end = var_9107_end_0, end_mask = var_9107_end_mask_0, x = read_state_1)[name = string("op_9107_cast_fp16")];
            tensor<int32, [4]> var_9117_begin_0 = const()[name = string("op_9117_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9117_end_0 = const()[name = string("op_9117_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9117_end_mask_0 = const()[name = string("op_9117_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9117_cast_fp16 = slice_by_index(begin = var_9117_begin_0, end = var_9117_end_0, end_mask = var_9117_end_mask_0, x = var_9107_cast_fp16)[name = string("op_9117_cast_fp16")];
            bool var_9123_transpose_x_0 = const()[name = string("op_9123_transpose_x_0"), val = bool(false)];
            bool var_9123_transpose_y_0 = const()[name = string("op_9123_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9123_cast_fp16 = matmul(transpose_x = var_9123_transpose_x_0, transpose_y = var_9123_transpose_y_0, x = var_9102_cast_fp16, y = var_9117_cast_fp16)[name = string("op_9123_cast_fp16")];
            tensor<int32, [4]> var_9128_begin_0 = const()[name = string("op_9128_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9128_end_0 = const()[name = string("op_9128_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9128_end_mask_0 = const()[name = string("op_9128_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9128_cast_fp16 = slice_by_index(begin = var_9128_begin_0, end = var_9128_end_0, end_mask = var_9128_end_mask_0, x = shares_51_cast_fp16)[name = string("op_9128_cast_fp16")];
            bool var_9129_transpose_x_0 = const()[name = string("op_9129_transpose_x_0"), val = bool(false)];
            bool var_9129_transpose_y_0 = const()[name = string("op_9129_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_51_cast_fp16 = transpose(perm = v_51_perm_0, x = var_8984_cast_fp16)[name = string("transpose_11")];
            tensor<fp16, [1, 8, 32, 128]> var_9129_cast_fp16 = matmul(transpose_x = var_9129_transpose_x_0, transpose_y = var_9129_transpose_y_0, x = var_9128_cast_fp16, y = v_51_cast_fp16)[name = string("op_9129_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_101_cast_fp16 = add(x = var_9123_cast_fp16, y = var_9129_cast_fp16)[name = string("mixed_101_cast_fp16")];
            tensor<int32, [4]> var_9146_begin_0 = const()[name = string("op_9146_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9146_end_0 = const()[name = string("op_9146_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9146_end_mask_0 = const()[name = string("op_9146_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9146_cast_fp16 = slice_by_index(begin = var_9146_begin_0, end = var_9146_end_0, end_mask = var_9146_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9146_cast_fp16")];
            tensor<int32, [4]> var_9161_begin_0 = const()[name = string("op_9161_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9161_end_0 = const()[name = string("op_9161_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9161_end_mask_0 = const()[name = string("op_9161_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9161_cast_fp16 = slice_by_index(begin = var_9161_begin_0, end = var_9161_end_0, end_mask = var_9161_end_mask_0, x = mixed_101_cast_fp16)[name = string("op_9161_cast_fp16")];
            int32 var_9163 = const()[name = string("op_9163"), val = int32(1)];
            bool mixed_103_interleave_0 = const()[name = string("mixed_103_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_103_cast_fp16 = concat(axis = var_9163, interleave = mixed_103_interleave_0, values = (var_9146_cast_fp16, var_9161_cast_fp16))[name = string("mixed_103_cast_fp16")];
            tensor<int32, [4]> var_9167_perm_0 = const()[name = string("op_9167_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9171 = const()[name = string("op_9171"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9167_cast_fp16 = transpose(perm = var_9167_perm_0, x = mixed_103_cast_fp16)[name = string("transpose_10")];
            tensor<fp16, [1, 16, 2048]> input_253_cast_fp16 = reshape(shape = var_9171, x = var_9167_cast_fp16)[name = string("input_253_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_25_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(709428672))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711525888))))[name = string("model_layers_25_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_178_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_self_attn_o_proj_weight_to_fp16_quantized, x = input_253_cast_fp16)[name = string("linear_178_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_311_cast_fp16 = add(x = x_301_cast_fp16, y = linear_178_cast_fp16)[name = string("x_311_cast_fp16")];
            fp16 var_9182_promoted_to_fp16 = const()[name = string("op_9182_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9184_cast_fp16 = pow(x = x_311_cast_fp16, y = var_9182_promoted_to_fp16)[name = string("op_9184_cast_fp16")];
            tensor<int32, [1]> var_9186_axes_0 = const()[name = string("op_9186_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9186_keep_dims_0 = const()[name = string("op_9186_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9186_cast_fp16 = reduce_mean(axes = var_9186_axes_0, keep_dims = var_9186_keep_dims_0, x = var_9184_cast_fp16)[name = string("op_9186_cast_fp16")];
            fp16 var_9187_to_fp16 = const()[name = string("op_9187_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9188_cast_fp16 = add(x = var_9186_cast_fp16, y = var_9187_to_fp16)[name = string("op_9188_cast_fp16")];
            fp32 var_9189_epsilon_0 = const()[name = string("op_9189_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9189_cast_fp16 = rsqrt(epsilon = var_9189_epsilon_0, x = var_9188_cast_fp16)[name = string("op_9189_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9190_cast_fp16 = mul(x = x_311_cast_fp16, y = var_9189_cast_fp16)[name = string("op_9190_cast_fp16")];
            tensor<fp16, [1024]> model_layers_25_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_25_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711528000)))];
            tensor<fp16, [1, 16, 1024]> input_255_cast_fp16 = mul(x = var_9190_cast_fp16, y = model_layers_25_post_attention_layernorm_weight_to_fp16)[name = string("input_255_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(711530112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714675904))))[name = string("model_layers_25_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_179_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_gate_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_179_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9195_cast_fp16 = silu(x = linear_179_cast_fp16)[name = string("op_9195_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_25_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(714682112))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717827904))))[name = string("model_layers_25_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_180_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_25_mlp_up_proj_weight_to_fp16_quantized, x = input_255_cast_fp16)[name = string("linear_180_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_259_cast_fp16 = mul(x = var_9195_cast_fp16, y = linear_180_cast_fp16)[name = string("input_259_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_25_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(717834112))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720979904))))[name = string("model_layers_25_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_181_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_25_mlp_down_proj_weight_to_fp16_quantized, x = input_259_cast_fp16)[name = string("linear_181_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_313_cast_fp16 = add(x = x_311_cast_fp16, y = linear_181_cast_fp16)[name = string("x_313_cast_fp16")];
            fp16 var_9209_promoted_to_fp16 = const()[name = string("op_9209_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9211_cast_fp16 = pow(x = x_313_cast_fp16, y = var_9209_promoted_to_fp16)[name = string("op_9211_cast_fp16")];
            tensor<int32, [1]> var_9213_axes_0 = const()[name = string("op_9213_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9213_keep_dims_0 = const()[name = string("op_9213_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9213_cast_fp16 = reduce_mean(axes = var_9213_axes_0, keep_dims = var_9213_keep_dims_0, x = var_9211_cast_fp16)[name = string("op_9213_cast_fp16")];
            fp16 var_9214_to_fp16 = const()[name = string("op_9214_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9215_cast_fp16 = add(x = var_9213_cast_fp16, y = var_9214_to_fp16)[name = string("op_9215_cast_fp16")];
            fp32 var_9216_epsilon_0 = const()[name = string("op_9216_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9216_cast_fp16 = rsqrt(epsilon = var_9216_epsilon_0, x = var_9215_cast_fp16)[name = string("op_9216_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9217_cast_fp16 = mul(x = x_313_cast_fp16, y = var_9216_cast_fp16)[name = string("op_9217_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720982016)))];
            tensor<fp16, [1, 16, 1024]> input_261_cast_fp16 = mul(x = var_9217_cast_fp16, y = model_layers_26_input_layernorm_weight_to_fp16)[name = string("input_261_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_26_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(720984128))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723081344))))[name = string("model_layers_26_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_182_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_26_self_attn_q_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_182_cast_fp16")];
            tensor<int32, [4]> var_9226 = const()[name = string("op_9226"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9227_cast_fp16 = reshape(shape = var_9226, x = linear_182_cast_fp16)[name = string("op_9227_cast_fp16")];
            tensor<int32, [4]> x_315_perm_0 = const()[name = string("x_315_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9235_promoted_to_fp16 = const()[name = string("op_9235_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_315_cast_fp16 = transpose(perm = x_315_perm_0, x = var_9227_cast_fp16)[name = string("transpose_9")];
            tensor<fp16, [1, 16, 16, 128]> var_9237_cast_fp16 = pow(x = x_315_cast_fp16, y = var_9235_promoted_to_fp16)[name = string("op_9237_cast_fp16")];
            tensor<int32, [1]> var_9239_axes_0 = const()[name = string("op_9239_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9239_keep_dims_0 = const()[name = string("op_9239_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9239_cast_fp16 = reduce_mean(axes = var_9239_axes_0, keep_dims = var_9239_keep_dims_0, x = var_9237_cast_fp16)[name = string("op_9239_cast_fp16")];
            fp16 var_9240_to_fp16 = const()[name = string("op_9240_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9241_cast_fp16 = add(x = var_9239_cast_fp16, y = var_9240_to_fp16)[name = string("op_9241_cast_fp16")];
            fp32 var_9242_epsilon_0 = const()[name = string("op_9242_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9242_cast_fp16 = rsqrt(epsilon = var_9242_epsilon_0, x = var_9241_cast_fp16)[name = string("op_9242_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9243_cast_fp16 = mul(x = x_315_cast_fp16, y = var_9242_cast_fp16)[name = string("op_9243_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085504)))];
            tensor<fp16, [1, 16, 16, 128]> x_319_cast_fp16 = mul(x = var_9243_cast_fp16, y = model_layers_26_self_attn_q_norm_weight_to_fp16)[name = string("x_319_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(723085824))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724134464))))[name = string("model_layers_26_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_183_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_k_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_183_cast_fp16")];
            tensor<int32, [4]> var_9252 = const()[name = string("op_9252"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9253_cast_fp16 = reshape(shape = var_9252, x = linear_183_cast_fp16)[name = string("op_9253_cast_fp16")];
            tensor<int32, [4]> x_317_perm_0 = const()[name = string("x_317_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9261_promoted_to_fp16 = const()[name = string("op_9261_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_317_cast_fp16 = transpose(perm = x_317_perm_0, x = var_9253_cast_fp16)[name = string("transpose_8")];
            tensor<fp16, [1, 8, 16, 128]> var_9263_cast_fp16 = pow(x = x_317_cast_fp16, y = var_9261_promoted_to_fp16)[name = string("op_9263_cast_fp16")];
            tensor<int32, [1]> var_9265_axes_0 = const()[name = string("op_9265_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9265_keep_dims_0 = const()[name = string("op_9265_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9265_cast_fp16 = reduce_mean(axes = var_9265_axes_0, keep_dims = var_9265_keep_dims_0, x = var_9263_cast_fp16)[name = string("op_9265_cast_fp16")];
            fp16 var_9266_to_fp16 = const()[name = string("op_9266_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9267_cast_fp16 = add(x = var_9265_cast_fp16, y = var_9266_to_fp16)[name = string("op_9267_cast_fp16")];
            fp32 var_9268_epsilon_0 = const()[name = string("op_9268_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9268_cast_fp16 = rsqrt(epsilon = var_9268_epsilon_0, x = var_9267_cast_fp16)[name = string("op_9268_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9269_cast_fp16 = mul(x = x_317_cast_fp16, y = var_9268_cast_fp16)[name = string("op_9269_cast_fp16")];
            tensor<fp16, [128]> model_layers_26_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_26_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136576)))];
            tensor<fp16, [1, 8, 16, 128]> x_321_cast_fp16 = mul(x = var_9269_cast_fp16, y = model_layers_26_self_attn_k_norm_weight_to_fp16)[name = string("x_321_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_26_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(724136896))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725185536))))[name = string("model_layers_26_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_184_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_v_proj_weight_to_fp16_quantized, x = input_261_cast_fp16)[name = string("linear_184_cast_fp16")];
            tensor<int32, [4]> var_9278 = const()[name = string("op_9278"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9279_cast_fp16 = reshape(shape = var_9278, x = linear_184_cast_fp16)[name = string("op_9279_cast_fp16")];
            tensor<int32, [4]> v_53_perm_0 = const()[name = string("v_53_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_105_begin_0 = const()[name = string("first_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_105_end_0 = const()[name = string("first_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_105_end_mask_0 = const()[name = string("first_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_105_cast_fp16 = slice_by_index(begin = first_105_begin_0, end = first_105_end_0, end_mask = first_105_end_mask_0, x = x_319_cast_fp16)[name = string("first_105_cast_fp16")];
            tensor<int32, [4]> second_105_begin_0 = const()[name = string("second_105_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_105_end_0 = const()[name = string("second_105_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_105_end_mask_0 = const()[name = string("second_105_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_105_cast_fp16 = slice_by_index(begin = second_105_begin_0, end = second_105_end_0, end_mask = second_105_end_mask_0, x = x_319_cast_fp16)[name = string("second_105_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9293_cast_fp16 = mul(x = first_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9293_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9294_cast_fp16 = mul(x = second_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9294_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9296_cast_fp16 = sub(x = var_9293_cast_fp16, y = var_9294_cast_fp16)[name = string("op_9296_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9297_cast_fp16 = mul(x = second_105_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9297_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9298_cast_fp16 = mul(x = first_105_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9298_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9300_cast_fp16 = add(x = var_9297_cast_fp16, y = var_9298_cast_fp16)[name = string("op_9300_cast_fp16")];
            int32 var_9302 = const()[name = string("op_9302"), val = int32(-1)];
            bool q_53_interleave_0 = const()[name = string("q_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_53_cast_fp16 = concat(axis = var_9302, interleave = q_53_interleave_0, values = (var_9296_cast_fp16, var_9300_cast_fp16))[name = string("q_53_cast_fp16")];
            tensor<int32, [4]> first_107_begin_0 = const()[name = string("first_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_107_end_0 = const()[name = string("first_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_107_end_mask_0 = const()[name = string("first_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_107_cast_fp16 = slice_by_index(begin = first_107_begin_0, end = first_107_end_0, end_mask = first_107_end_mask_0, x = x_321_cast_fp16)[name = string("first_107_cast_fp16")];
            tensor<int32, [4]> second_107_begin_0 = const()[name = string("second_107_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_107_end_0 = const()[name = string("second_107_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_107_end_mask_0 = const()[name = string("second_107_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_107_cast_fp16 = slice_by_index(begin = second_107_begin_0, end = second_107_end_0, end_mask = second_107_end_mask_0, x = x_321_cast_fp16)[name = string("second_107_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9314_cast_fp16 = mul(x = first_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9314_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9315_cast_fp16 = mul(x = second_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9315_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9317_cast_fp16 = sub(x = var_9314_cast_fp16, y = var_9315_cast_fp16)[name = string("op_9317_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9318_cast_fp16 = mul(x = second_107_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9318_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9319_cast_fp16 = mul(x = first_107_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9319_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9321_cast_fp16 = add(x = var_9318_cast_fp16, y = var_9319_cast_fp16)[name = string("op_9321_cast_fp16")];
            int32 var_9323 = const()[name = string("op_9323"), val = int32(-1)];
            bool k_53_interleave_0 = const()[name = string("k_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_53_cast_fp16 = concat(axis = var_9323, interleave = k_53_interleave_0, values = (var_9317_cast_fp16, var_9321_cast_fp16))[name = string("k_53_cast_fp16")];
            tensor<int32, [4]> var_9334_begin_0 = const()[name = string("op_9334_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9334_end_0 = const()[name = string("op_9334_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9334_end_mask_0 = const()[name = string("op_9334_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9334_stride_0 = const()[name = string("op_9334_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9334_cast_fp16 = slice_by_index(begin = var_9334_begin_0, end = var_9334_end_0, end_mask = var_9334_end_mask_0, stride = var_9334_stride_0, x = q_53_cast_fp16)[name = string("op_9334_cast_fp16")];
            tensor<int32, [4]> var_9344_begin_0 = const()[name = string("op_9344_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9344_end_0 = const()[name = string("op_9344_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9344_end_mask_0 = const()[name = string("op_9344_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9344_stride_0 = const()[name = string("op_9344_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9344_cast_fp16 = slice_by_index(begin = var_9344_begin_0, end = var_9344_end_0, end_mask = var_9344_end_mask_0, stride = var_9344_stride_0, x = q_53_cast_fp16)[name = string("op_9344_cast_fp16")];
            int32 var_9346 = const()[name = string("op_9346"), val = int32(2)];
            bool stacked_53_interleave_0 = const()[name = string("stacked_53_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_53_cast_fp16 = concat(axis = var_9346, interleave = stacked_53_interleave_0, values = (var_9334_cast_fp16, var_9344_cast_fp16))[name = string("stacked_53_cast_fp16")];
            tensor<int32, [4]> var_9352_begin_0 = const()[name = string("op_9352_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9352_end_0 = const()[name = string("op_9352_end_0"), val = tensor<int32, [4]>([27, 8, 528, 128])];
            tensor<bool, [4]> var_9352_end_mask_0 = const()[name = string("op_9352_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9352_cast_fp16 = slice_by_index(begin = var_9352_begin_0, end = var_9352_end_0, end_mask = var_9352_end_mask_0, x = read_state_0)[name = string("op_9352_cast_fp16")];
            tensor<int32, [4]> var_9362_begin_0 = const()[name = string("op_9362_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9362_end_0 = const()[name = string("op_9362_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9362_end_mask_0 = const()[name = string("op_9362_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9362_cast_fp16 = slice_by_index(begin = var_9362_begin_0, end = var_9362_end_0, end_mask = var_9362_end_mask_0, x = var_9352_cast_fp16)[name = string("op_9362_cast_fp16")];
            bool var_9371_transpose_x_1 = const()[name = string("op_9371_transpose_x_1"), val = bool(false)];
            bool var_9371_transpose_y_1 = const()[name = string("op_9371_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9371_cast_fp16 = matmul(transpose_x = var_9371_transpose_x_1, transpose_y = var_9371_transpose_y_1, x = var_9362_cast_fp16, y = stacked_53_cast_fp16)[name = string("op_9371_cast_fp16")];
            tensor<int32, [4]> var_9374_perm_0 = const()[name = string("op_9374_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9375_to_fp16 = const()[name = string("op_9375_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9374_cast_fp16 = transpose(perm = var_9374_perm_0, x = var_9371_cast_fp16)[name = string("transpose_7")];
            tensor<fp16, [1, 8, 32, 512]> var_9376_cast_fp16 = mul(x = var_9374_cast_fp16, y = var_9375_to_fp16)[name = string("op_9376_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_53_cast_fp16 = add(x = var_9376_cast_fp16, y = cacheMask)[name = string("cache_weights_53_cast_fp16")];
            bool var_9382_transpose_x_1 = const()[name = string("op_9382_transpose_x_1"), val = bool(false)];
            bool var_9382_transpose_y_1 = const()[name = string("op_9382_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9382_cast_fp16 = matmul(transpose_x = var_9382_transpose_x_1, transpose_y = var_9382_transpose_y_1, x = stacked_53_cast_fp16, y = k_53_cast_fp16)[name = string("op_9382_cast_fp16")];
            fp16 var_9383_to_fp16 = const()[name = string("op_9383_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9384_cast_fp16 = mul(x = var_9382_cast_fp16, y = var_9383_to_fp16)[name = string("op_9384_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_53_cast_fp16 = add(x = var_9384_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_53_cast_fp16")];
            int32 var_9388 = const()[name = string("op_9388"), val = int32(-1)];
            bool var_9389_interleave_0 = const()[name = string("op_9389_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9389_cast_fp16 = concat(axis = var_9388, interleave = var_9389_interleave_0, values = (cache_weights_53_cast_fp16, call_weights_53_cast_fp16))[name = string("op_9389_cast_fp16")];
            int32 var_9390 = const()[name = string("op_9390"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_53_cast_fp16 = softmax(axis = var_9390, x = var_9389_cast_fp16)[name = string("shares_53_cast_fp16")];
            tensor<int32, [4]> var_9397_begin_0 = const()[name = string("op_9397_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9397_end_0 = const()[name = string("op_9397_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9397_end_mask_0 = const()[name = string("op_9397_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9397_cast_fp16 = slice_by_index(begin = var_9397_begin_0, end = var_9397_end_0, end_mask = var_9397_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9397_cast_fp16")];
            tensor<int32, [4]> var_9402_begin_0 = const()[name = string("op_9402_begin_0"), val = tensor<int32, [4]>([26, 0, 0, 0])];
            tensor<int32, [4]> var_9402_end_0 = const()[name = string("op_9402_end_0"), val = tensor<int32, [4]>([27, 8, 528, 128])];
            tensor<bool, [4]> var_9402_end_mask_0 = const()[name = string("op_9402_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9402_cast_fp16 = slice_by_index(begin = var_9402_begin_0, end = var_9402_end_0, end_mask = var_9402_end_mask_0, x = read_state_1)[name = string("op_9402_cast_fp16")];
            tensor<int32, [4]> var_9412_begin_0 = const()[name = string("op_9412_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9412_end_0 = const()[name = string("op_9412_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9412_end_mask_0 = const()[name = string("op_9412_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9412_cast_fp16 = slice_by_index(begin = var_9412_begin_0, end = var_9412_end_0, end_mask = var_9412_end_mask_0, x = var_9402_cast_fp16)[name = string("op_9412_cast_fp16")];
            bool var_9418_transpose_x_0 = const()[name = string("op_9418_transpose_x_0"), val = bool(false)];
            bool var_9418_transpose_y_0 = const()[name = string("op_9418_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9418_cast_fp16 = matmul(transpose_x = var_9418_transpose_x_0, transpose_y = var_9418_transpose_y_0, x = var_9397_cast_fp16, y = var_9412_cast_fp16)[name = string("op_9418_cast_fp16")];
            tensor<int32, [4]> var_9423_begin_0 = const()[name = string("op_9423_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9423_end_0 = const()[name = string("op_9423_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9423_end_mask_0 = const()[name = string("op_9423_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9423_cast_fp16 = slice_by_index(begin = var_9423_begin_0, end = var_9423_end_0, end_mask = var_9423_end_mask_0, x = shares_53_cast_fp16)[name = string("op_9423_cast_fp16")];
            bool var_9424_transpose_x_0 = const()[name = string("op_9424_transpose_x_0"), val = bool(false)];
            bool var_9424_transpose_y_0 = const()[name = string("op_9424_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_53_cast_fp16 = transpose(perm = v_53_perm_0, x = var_9279_cast_fp16)[name = string("transpose_6")];
            tensor<fp16, [1, 8, 32, 128]> var_9424_cast_fp16 = matmul(transpose_x = var_9424_transpose_x_0, transpose_y = var_9424_transpose_y_0, x = var_9423_cast_fp16, y = v_53_cast_fp16)[name = string("op_9424_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_105_cast_fp16 = add(x = var_9418_cast_fp16, y = var_9424_cast_fp16)[name = string("mixed_105_cast_fp16")];
            tensor<int32, [4]> var_9441_begin_0 = const()[name = string("op_9441_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9441_end_0 = const()[name = string("op_9441_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9441_end_mask_0 = const()[name = string("op_9441_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9441_cast_fp16 = slice_by_index(begin = var_9441_begin_0, end = var_9441_end_0, end_mask = var_9441_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9441_cast_fp16")];
            tensor<int32, [4]> var_9456_begin_0 = const()[name = string("op_9456_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9456_end_0 = const()[name = string("op_9456_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9456_end_mask_0 = const()[name = string("op_9456_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9456_cast_fp16 = slice_by_index(begin = var_9456_begin_0, end = var_9456_end_0, end_mask = var_9456_end_mask_0, x = mixed_105_cast_fp16)[name = string("op_9456_cast_fp16")];
            int32 var_9458 = const()[name = string("op_9458"), val = int32(1)];
            bool mixed_107_interleave_0 = const()[name = string("mixed_107_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_107_cast_fp16 = concat(axis = var_9458, interleave = mixed_107_interleave_0, values = (var_9441_cast_fp16, var_9456_cast_fp16))[name = string("mixed_107_cast_fp16")];
            tensor<int32, [4]> var_9462_perm_0 = const()[name = string("op_9462_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9466 = const()[name = string("op_9466"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9462_cast_fp16 = transpose(perm = var_9462_perm_0, x = mixed_107_cast_fp16)[name = string("transpose_5")];
            tensor<fp16, [1, 16, 2048]> input_263_cast_fp16 = reshape(shape = var_9466, x = var_9462_cast_fp16)[name = string("input_263_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_26_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(725187648))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727284864))))[name = string("model_layers_26_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_185_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_self_attn_o_proj_weight_to_fp16_quantized, x = input_263_cast_fp16)[name = string("linear_185_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_323_cast_fp16 = add(x = x_313_cast_fp16, y = linear_185_cast_fp16)[name = string("x_323_cast_fp16")];
            fp16 var_9477_promoted_to_fp16 = const()[name = string("op_9477_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9479_cast_fp16 = pow(x = x_323_cast_fp16, y = var_9477_promoted_to_fp16)[name = string("op_9479_cast_fp16")];
            tensor<int32, [1]> var_9481_axes_0 = const()[name = string("op_9481_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9481_keep_dims_0 = const()[name = string("op_9481_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9481_cast_fp16 = reduce_mean(axes = var_9481_axes_0, keep_dims = var_9481_keep_dims_0, x = var_9479_cast_fp16)[name = string("op_9481_cast_fp16")];
            fp16 var_9482_to_fp16 = const()[name = string("op_9482_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9483_cast_fp16 = add(x = var_9481_cast_fp16, y = var_9482_to_fp16)[name = string("op_9483_cast_fp16")];
            fp32 var_9484_epsilon_0 = const()[name = string("op_9484_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9484_cast_fp16 = rsqrt(epsilon = var_9484_epsilon_0, x = var_9483_cast_fp16)[name = string("op_9484_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9485_cast_fp16 = mul(x = x_323_cast_fp16, y = var_9484_cast_fp16)[name = string("op_9485_cast_fp16")];
            tensor<fp16, [1024]> model_layers_26_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_26_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727286976)))];
            tensor<fp16, [1, 16, 1024]> input_265_cast_fp16 = mul(x = var_9485_cast_fp16, y = model_layers_26_post_attention_layernorm_weight_to_fp16)[name = string("input_265_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(727289088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730434880))))[name = string("model_layers_26_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_186_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_gate_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_186_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9490_cast_fp16 = silu(x = linear_186_cast_fp16)[name = string("op_9490_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_26_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(730441088))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733586880))))[name = string("model_layers_26_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_187_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_26_mlp_up_proj_weight_to_fp16_quantized, x = input_265_cast_fp16)[name = string("linear_187_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_269_cast_fp16 = mul(x = var_9490_cast_fp16, y = linear_187_cast_fp16)[name = string("input_269_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_26_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(733593088))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736738880))))[name = string("model_layers_26_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_188_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_26_mlp_down_proj_weight_to_fp16_quantized, x = input_269_cast_fp16)[name = string("linear_188_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_325_cast_fp16 = add(x = x_323_cast_fp16, y = linear_188_cast_fp16)[name = string("x_325_cast_fp16")];
            fp16 var_9504_promoted_to_fp16 = const()[name = string("op_9504_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9506_cast_fp16 = pow(x = x_325_cast_fp16, y = var_9504_promoted_to_fp16)[name = string("op_9506_cast_fp16")];
            tensor<int32, [1]> var_9508_axes_0 = const()[name = string("op_9508_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9508_keep_dims_0 = const()[name = string("op_9508_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9508_cast_fp16 = reduce_mean(axes = var_9508_axes_0, keep_dims = var_9508_keep_dims_0, x = var_9506_cast_fp16)[name = string("op_9508_cast_fp16")];
            fp16 var_9509_to_fp16 = const()[name = string("op_9509_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9510_cast_fp16 = add(x = var_9508_cast_fp16, y = var_9509_to_fp16)[name = string("op_9510_cast_fp16")];
            fp32 var_9511_epsilon_0 = const()[name = string("op_9511_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9511_cast_fp16 = rsqrt(epsilon = var_9511_epsilon_0, x = var_9510_cast_fp16)[name = string("op_9511_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9512_cast_fp16 = mul(x = x_325_cast_fp16, y = var_9511_cast_fp16)[name = string("op_9512_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_input_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_input_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736740992)))];
            tensor<fp16, [1, 16, 1024]> input_271_cast_fp16 = mul(x = var_9512_cast_fp16, y = model_layers_27_input_layernorm_weight_to_fp16)[name = string("input_271_cast_fp16")];
            tensor<fp16, [2048, 1024]> model_layers_27_self_attn_q_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [2048, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(736743104))), scale = tensor<fp16, [2048, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738840320))))[name = string("model_layers_27_self_attn_q_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 2048]> linear_189_cast_fp16 = linear(bias = linear_0_bias_0_to_fp16, weight = model_layers_27_self_attn_q_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_189_cast_fp16")];
            tensor<int32, [4]> var_9521 = const()[name = string("op_9521"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<fp16, [1, 16, 16, 128]> var_9522_cast_fp16 = reshape(shape = var_9521, x = linear_189_cast_fp16)[name = string("op_9522_cast_fp16")];
            tensor<int32, [4]> x_327_perm_0 = const()[name = string("x_327_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9530_promoted_to_fp16 = const()[name = string("op_9530_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 16, 128]> x_327_cast_fp16 = transpose(perm = x_327_perm_0, x = var_9522_cast_fp16)[name = string("transpose_4")];
            tensor<fp16, [1, 16, 16, 128]> var_9532_cast_fp16 = pow(x = x_327_cast_fp16, y = var_9530_promoted_to_fp16)[name = string("op_9532_cast_fp16")];
            tensor<int32, [1]> var_9534_axes_0 = const()[name = string("op_9534_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9534_keep_dims_0 = const()[name = string("op_9534_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 16, 1]> var_9534_cast_fp16 = reduce_mean(axes = var_9534_axes_0, keep_dims = var_9534_keep_dims_0, x = var_9532_cast_fp16)[name = string("op_9534_cast_fp16")];
            fp16 var_9535_to_fp16 = const()[name = string("op_9535_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 16, 1]> var_9536_cast_fp16 = add(x = var_9534_cast_fp16, y = var_9535_to_fp16)[name = string("op_9536_cast_fp16")];
            fp32 var_9537_epsilon_0 = const()[name = string("op_9537_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 16, 1]> var_9537_cast_fp16 = rsqrt(epsilon = var_9537_epsilon_0, x = var_9536_cast_fp16)[name = string("op_9537_cast_fp16")];
            tensor<fp16, [1, 16, 16, 128]> var_9538_cast_fp16 = mul(x = x_327_cast_fp16, y = var_9537_cast_fp16)[name = string("op_9538_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_q_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_q_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844480)))];
            tensor<fp16, [1, 16, 16, 128]> x_331_cast_fp16 = mul(x = var_9538_cast_fp16, y = model_layers_27_self_attn_q_norm_weight_to_fp16)[name = string("x_331_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_k_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(738844800))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739893440))))[name = string("model_layers_27_self_attn_k_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_190_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_k_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_190_cast_fp16")];
            tensor<int32, [4]> var_9547 = const()[name = string("op_9547"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9548_cast_fp16 = reshape(shape = var_9547, x = linear_190_cast_fp16)[name = string("op_9548_cast_fp16")];
            tensor<int32, [4]> x_329_perm_0 = const()[name = string("x_329_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            fp16 var_9556_promoted_to_fp16 = const()[name = string("op_9556_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 8, 16, 128]> x_329_cast_fp16 = transpose(perm = x_329_perm_0, x = var_9548_cast_fp16)[name = string("transpose_3")];
            tensor<fp16, [1, 8, 16, 128]> var_9558_cast_fp16 = pow(x = x_329_cast_fp16, y = var_9556_promoted_to_fp16)[name = string("op_9558_cast_fp16")];
            tensor<int32, [1]> var_9560_axes_0 = const()[name = string("op_9560_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9560_keep_dims_0 = const()[name = string("op_9560_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 8, 16, 1]> var_9560_cast_fp16 = reduce_mean(axes = var_9560_axes_0, keep_dims = var_9560_keep_dims_0, x = var_9558_cast_fp16)[name = string("op_9560_cast_fp16")];
            fp16 var_9561_to_fp16 = const()[name = string("op_9561_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 8, 16, 1]> var_9562_cast_fp16 = add(x = var_9560_cast_fp16, y = var_9561_to_fp16)[name = string("op_9562_cast_fp16")];
            fp32 var_9563_epsilon_0 = const()[name = string("op_9563_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 8, 16, 1]> var_9563_cast_fp16 = rsqrt(epsilon = var_9563_epsilon_0, x = var_9562_cast_fp16)[name = string("op_9563_cast_fp16")];
            tensor<fp16, [1, 8, 16, 128]> var_9564_cast_fp16 = mul(x = x_329_cast_fp16, y = var_9563_cast_fp16)[name = string("op_9564_cast_fp16")];
            tensor<fp16, [128]> model_layers_27_self_attn_k_norm_weight_to_fp16 = const()[name = string("model_layers_27_self_attn_k_norm_weight_to_fp16"), val = tensor<fp16, [128]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895552)))];
            tensor<fp16, [1, 8, 16, 128]> x_333_cast_fp16 = mul(x = var_9564_cast_fp16, y = model_layers_27_self_attn_k_norm_weight_to_fp16)[name = string("x_333_cast_fp16")];
            tensor<fp16, [1024, 1024]> model_layers_27_self_attn_v_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(739895872))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740944512))))[name = string("model_layers_27_self_attn_v_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_191_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_v_proj_weight_to_fp16_quantized, x = input_271_cast_fp16)[name = string("linear_191_cast_fp16")];
            tensor<int32, [4]> var_9573 = const()[name = string("op_9573"), val = tensor<int32, [4]>([1, 16, 8, 128])];
            tensor<fp16, [1, 16, 8, 128]> var_9574_cast_fp16 = reshape(shape = var_9573, x = linear_191_cast_fp16)[name = string("op_9574_cast_fp16")];
            tensor<int32, [4]> v_perm_0 = const()[name = string("v_perm_0"), val = tensor<int32, [4]>([0, 2, -3, -1])];
            tensor<int32, [4]> first_109_begin_0 = const()[name = string("first_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_109_end_0 = const()[name = string("first_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 64])];
            tensor<bool, [4]> first_109_end_mask_0 = const()[name = string("first_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 16, 16, 64]> first_109_cast_fp16 = slice_by_index(begin = first_109_begin_0, end = first_109_end_0, end_mask = first_109_end_mask_0, x = x_331_cast_fp16)[name = string("first_109_cast_fp16")];
            tensor<int32, [4]> second_109_begin_0 = const()[name = string("second_109_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_109_end_0 = const()[name = string("second_109_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> second_109_end_mask_0 = const()[name = string("second_109_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 16, 16, 64]> second_109_cast_fp16 = slice_by_index(begin = second_109_begin_0, end = second_109_end_0, end_mask = second_109_end_mask_0, x = x_331_cast_fp16)[name = string("second_109_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9588_cast_fp16 = mul(x = first_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9588_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9589_cast_fp16 = mul(x = second_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9589_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9591_cast_fp16 = sub(x = var_9588_cast_fp16, y = var_9589_cast_fp16)[name = string("op_9591_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9592_cast_fp16 = mul(x = second_109_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9592_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9593_cast_fp16 = mul(x = first_109_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9593_cast_fp16")];
            tensor<fp16, [1, 16, 16, 64]> var_9595_cast_fp16 = add(x = var_9592_cast_fp16, y = var_9593_cast_fp16)[name = string("op_9595_cast_fp16")];
            int32 var_9597 = const()[name = string("op_9597"), val = int32(-1)];
            bool q_interleave_0 = const()[name = string("q_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> q_cast_fp16 = concat(axis = var_9597, interleave = q_interleave_0, values = (var_9591_cast_fp16, var_9595_cast_fp16))[name = string("q_cast_fp16")];
            tensor<int32, [4]> first_begin_0 = const()[name = string("first_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> first_end_0 = const()[name = string("first_end_0"), val = tensor<int32, [4]>([1, 8, 16, 64])];
            tensor<bool, [4]> first_end_mask_0 = const()[name = string("first_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 16, 64]> first_cast_fp16 = slice_by_index(begin = first_begin_0, end = first_end_0, end_mask = first_end_mask_0, x = x_333_cast_fp16)[name = string("first_cast_fp16")];
            tensor<int32, [4]> second_begin_0 = const()[name = string("second_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 64])];
            tensor<int32, [4]> second_end_0 = const()[name = string("second_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> second_end_mask_0 = const()[name = string("second_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 64]> second_cast_fp16 = slice_by_index(begin = second_begin_0, end = second_end_0, end_mask = second_end_mask_0, x = x_333_cast_fp16)[name = string("second_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9609_cast_fp16 = mul(x = first_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9609_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9610_cast_fp16 = mul(x = second_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9610_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9612_cast_fp16 = sub(x = var_9609_cast_fp16, y = var_9610_cast_fp16)[name = string("op_9612_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9613_cast_fp16 = mul(x = second_cast_fp16, y = cos_1_cast_fp16)[name = string("op_9613_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9614_cast_fp16 = mul(x = first_cast_fp16, y = sin_1_cast_fp16)[name = string("op_9614_cast_fp16")];
            tensor<fp16, [1, 8, 16, 64]> var_9616_cast_fp16 = add(x = var_9613_cast_fp16, y = var_9614_cast_fp16)[name = string("op_9616_cast_fp16")];
            int32 var_9618 = const()[name = string("op_9618"), val = int32(-1)];
            bool k_interleave_0 = const()[name = string("k_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> k_cast_fp16 = concat(axis = var_9618, interleave = k_interleave_0, values = (var_9612_cast_fp16, var_9616_cast_fp16))[name = string("k_cast_fp16")];
            tensor<int32, [4]> var_9629_begin_0 = const()[name = string("op_9629_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9629_end_0 = const()[name = string("op_9629_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9629_end_mask_0 = const()[name = string("op_9629_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9629_stride_0 = const()[name = string("op_9629_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9629_cast_fp16 = slice_by_index(begin = var_9629_begin_0, end = var_9629_end_0, end_mask = var_9629_end_mask_0, stride = var_9629_stride_0, x = q_cast_fp16)[name = string("op_9629_cast_fp16")];
            tensor<int32, [4]> var_9639_begin_0 = const()[name = string("op_9639_begin_0"), val = tensor<int32, [4]>([0, 1, 0, 0])];
            tensor<int32, [4]> var_9639_end_0 = const()[name = string("op_9639_end_0"), val = tensor<int32, [4]>([1, 16, 16, 128])];
            tensor<bool, [4]> var_9639_end_mask_0 = const()[name = string("op_9639_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<int32, [4]> var_9639_stride_0 = const()[name = string("op_9639_stride_0"), val = tensor<int32, [4]>([1, 2, 1, 1])];
            tensor<fp16, [1, 8, 16, 128]> var_9639_cast_fp16 = slice_by_index(begin = var_9639_begin_0, end = var_9639_end_0, end_mask = var_9639_end_mask_0, stride = var_9639_stride_0, x = q_cast_fp16)[name = string("op_9639_cast_fp16")];
            int32 var_9641 = const()[name = string("op_9641"), val = int32(2)];
            bool stacked_interleave_0 = const()[name = string("stacked_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> stacked_cast_fp16 = concat(axis = var_9641, interleave = stacked_interleave_0, values = (var_9629_cast_fp16, var_9639_cast_fp16))[name = string("stacked_cast_fp16")];
            tensor<int32, [4]> var_9647_begin_0 = const()[name = string("op_9647_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9647_end_0 = const()[name = string("op_9647_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_9647_end_mask_0 = const()[name = string("op_9647_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9647_cast_fp16 = slice_by_index(begin = var_9647_begin_0, end = var_9647_end_0, end_mask = var_9647_end_mask_0, x = read_state_0)[name = string("op_9647_cast_fp16")];
            tensor<int32, [4]> var_9657_begin_0 = const()[name = string("op_9657_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9657_end_0 = const()[name = string("op_9657_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9657_end_mask_0 = const()[name = string("op_9657_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9657_cast_fp16 = slice_by_index(begin = var_9657_begin_0, end = var_9657_end_0, end_mask = var_9657_end_mask_0, x = var_9647_cast_fp16)[name = string("op_9657_cast_fp16")];
            bool var_9666_transpose_x_1 = const()[name = string("op_9666_transpose_x_1"), val = bool(false)];
            bool var_9666_transpose_y_1 = const()[name = string("op_9666_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 512, 32]> var_9666_cast_fp16 = matmul(transpose_x = var_9666_transpose_x_1, transpose_y = var_9666_transpose_y_1, x = var_9657_cast_fp16, y = stacked_cast_fp16)[name = string("op_9666_cast_fp16")];
            tensor<int32, [4]> var_9669_perm_0 = const()[name = string("op_9669_perm_0"), val = tensor<int32, [4]>([0, 1, -1, -2])];
            fp16 var_9670_to_fp16 = const()[name = string("op_9670_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 512]> var_9669_cast_fp16 = transpose(perm = var_9669_perm_0, x = var_9666_cast_fp16)[name = string("transpose_2")];
            tensor<fp16, [1, 8, 32, 512]> var_9671_cast_fp16 = mul(x = var_9669_cast_fp16, y = var_9670_to_fp16)[name = string("op_9671_cast_fp16")];
            tensor<fp16, [1, 8, 32, 512]> cache_weights_cast_fp16 = add(x = var_9671_cast_fp16, y = cacheMask)[name = string("cache_weights_cast_fp16")];
            bool var_9677_transpose_x_1 = const()[name = string("op_9677_transpose_x_1"), val = bool(false)];
            bool var_9677_transpose_y_1 = const()[name = string("op_9677_transpose_y_1"), val = bool(true)];
            tensor<fp16, [1, 8, 32, 16]> var_9677_cast_fp16 = matmul(transpose_x = var_9677_transpose_x_1, transpose_y = var_9677_transpose_y_1, x = stacked_cast_fp16, y = k_cast_fp16)[name = string("op_9677_cast_fp16")];
            fp16 var_9678_to_fp16 = const()[name = string("op_9678_to_fp16"), val = fp16(0x1.6ap-4)];
            tensor<fp16, [1, 8, 32, 16]> var_9679_cast_fp16 = mul(x = var_9677_cast_fp16, y = var_9678_to_fp16)[name = string("op_9679_cast_fp16")];
            tensor<fp16, [1, 8, 32, 16]> call_weights_cast_fp16 = add(x = var_9679_cast_fp16, y = later_rows_mask_to_fp16)[name = string("call_weights_cast_fp16")];
            int32 var_9683 = const()[name = string("op_9683"), val = int32(-1)];
            bool var_9684_interleave_0 = const()[name = string("op_9684_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 528]> var_9684_cast_fp16 = concat(axis = var_9683, interleave = var_9684_interleave_0, values = (cache_weights_cast_fp16, call_weights_cast_fp16))[name = string("op_9684_cast_fp16")];
            int32 var_9685 = const()[name = string("op_9685"), val = int32(-1)];
            tensor<fp16, [1, 8, 32, 528]> shares_cast_fp16 = softmax(axis = var_9685, x = var_9684_cast_fp16)[name = string("shares_cast_fp16")];
            tensor<int32, [4]> var_9692_begin_0 = const()[name = string("op_9692_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9692_end_0 = const()[name = string("op_9692_end_0"), val = tensor<int32, [4]>([1, 8, 32, 512])];
            tensor<bool, [4]> var_9692_end_mask_0 = const()[name = string("op_9692_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 8, 32, 512]> var_9692_cast_fp16 = slice_by_index(begin = var_9692_begin_0, end = var_9692_end_0, end_mask = var_9692_end_mask_0, x = shares_cast_fp16)[name = string("op_9692_cast_fp16")];
            tensor<int32, [4]> var_9697_begin_0 = const()[name = string("op_9697_begin_0"), val = tensor<int32, [4]>([27, 0, 0, 0])];
            tensor<int32, [4]> var_9697_end_0 = const()[name = string("op_9697_end_0"), val = tensor<int32, [4]>([1, 8, 528, 128])];
            tensor<bool, [4]> var_9697_end_mask_0 = const()[name = string("op_9697_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 528, 128]> var_9697_cast_fp16 = slice_by_index(begin = var_9697_begin_0, end = var_9697_end_0, end_mask = var_9697_end_mask_0, x = read_state_1)[name = string("op_9697_cast_fp16")];
            tensor<int32, [4]> var_9707_begin_0 = const()[name = string("op_9707_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9707_end_0 = const()[name = string("op_9707_end_0"), val = tensor<int32, [4]>([1, 8, 512, 128])];
            tensor<bool, [4]> var_9707_end_mask_0 = const()[name = string("op_9707_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 512, 128]> var_9707_cast_fp16 = slice_by_index(begin = var_9707_begin_0, end = var_9707_end_0, end_mask = var_9707_end_mask_0, x = var_9697_cast_fp16)[name = string("op_9707_cast_fp16")];
            bool var_9713_transpose_x_0 = const()[name = string("op_9713_transpose_x_0"), val = bool(false)];
            bool var_9713_transpose_y_0 = const()[name = string("op_9713_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 32, 128]> var_9713_cast_fp16 = matmul(transpose_x = var_9713_transpose_x_0, transpose_y = var_9713_transpose_y_0, x = var_9692_cast_fp16, y = var_9707_cast_fp16)[name = string("op_9713_cast_fp16")];
            tensor<int32, [4]> var_9718_begin_0 = const()[name = string("op_9718_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 512])];
            tensor<int32, [4]> var_9718_end_0 = const()[name = string("op_9718_end_0"), val = tensor<int32, [4]>([1, 8, 32, 528])];
            tensor<bool, [4]> var_9718_end_mask_0 = const()[name = string("op_9718_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 32, 16]> var_9718_cast_fp16 = slice_by_index(begin = var_9718_begin_0, end = var_9718_end_0, end_mask = var_9718_end_mask_0, x = shares_cast_fp16)[name = string("op_9718_cast_fp16")];
            bool var_9719_transpose_x_0 = const()[name = string("op_9719_transpose_x_0"), val = bool(false)];
            bool var_9719_transpose_y_0 = const()[name = string("op_9719_transpose_y_0"), val = bool(false)];
            tensor<fp16, [1, 8, 16, 128]> v_cast_fp16 = transpose(perm = v_perm_0, x = var_9574_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [1, 8, 32, 128]> var_9719_cast_fp16 = matmul(transpose_x = var_9719_transpose_x_0, transpose_y = var_9719_transpose_y_0, x = var_9718_cast_fp16, y = v_cast_fp16)[name = string("op_9719_cast_fp16")];
            tensor<fp16, [1, 8, 32, 128]> mixed_109_cast_fp16 = add(x = var_9713_cast_fp16, y = var_9719_cast_fp16)[name = string("mixed_109_cast_fp16")];
            tensor<int32, [4]> var_9736_begin_0 = const()[name = string("op_9736_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_9736_end_0 = const()[name = string("op_9736_end_0"), val = tensor<int32, [4]>([1, 8, 16, 128])];
            tensor<bool, [4]> var_9736_end_mask_0 = const()[name = string("op_9736_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9736_cast_fp16 = slice_by_index(begin = var_9736_begin_0, end = var_9736_end_0, end_mask = var_9736_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9736_cast_fp16")];
            tensor<int32, [4]> var_9751_begin_0 = const()[name = string("op_9751_begin_0"), val = tensor<int32, [4]>([0, 0, 16, 0])];
            tensor<int32, [4]> var_9751_end_0 = const()[name = string("op_9751_end_0"), val = tensor<int32, [4]>([1, 8, 1, 128])];
            tensor<bool, [4]> var_9751_end_mask_0 = const()[name = string("op_9751_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 8, 16, 128]> var_9751_cast_fp16 = slice_by_index(begin = var_9751_begin_0, end = var_9751_end_0, end_mask = var_9751_end_mask_0, x = mixed_109_cast_fp16)[name = string("op_9751_cast_fp16")];
            int32 var_9753 = const()[name = string("op_9753"), val = int32(1)];
            bool mixed_interleave_0 = const()[name = string("mixed_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 16, 16, 128]> mixed_cast_fp16 = concat(axis = var_9753, interleave = mixed_interleave_0, values = (var_9736_cast_fp16, var_9751_cast_fp16))[name = string("mixed_cast_fp16")];
            tensor<int32, [4]> var_9757_perm_0 = const()[name = string("op_9757_perm_0"), val = tensor<int32, [4]>([0, 2, 1, 3])];
            tensor<int32, [3]> var_9761 = const()[name = string("op_9761"), val = tensor<int32, [3]>([1, 16, 2048])];
            tensor<fp16, [1, 16, 16, 128]> var_9757_cast_fp16 = transpose(perm = var_9757_perm_0, x = mixed_cast_fp16)[name = string("transpose_0")];
            tensor<fp16, [1, 16, 2048]> input_273_cast_fp16 = reshape(shape = var_9761, x = var_9757_cast_fp16)[name = string("input_273_cast_fp16")];
            tensor<fp16, [1024, 2048]> model_layers_27_self_attn_o_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 2048]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(740946624))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743043840))))[name = string("model_layers_27_self_attn_o_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_192_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_self_attn_o_proj_weight_to_fp16_quantized, x = input_273_cast_fp16)[name = string("linear_192_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_335_cast_fp16 = add(x = x_325_cast_fp16, y = linear_192_cast_fp16)[name = string("x_335_cast_fp16")];
            fp16 var_9772_promoted_to_fp16 = const()[name = string("op_9772_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 16, 1024]> var_9774_cast_fp16 = pow(x = x_335_cast_fp16, y = var_9772_promoted_to_fp16)[name = string("op_9774_cast_fp16")];
            tensor<int32, [1]> var_9776_axes_0 = const()[name = string("op_9776_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9776_keep_dims_0 = const()[name = string("op_9776_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 16, 1]> var_9776_cast_fp16 = reduce_mean(axes = var_9776_axes_0, keep_dims = var_9776_keep_dims_0, x = var_9774_cast_fp16)[name = string("op_9776_cast_fp16")];
            fp16 var_9777_to_fp16 = const()[name = string("op_9777_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 16, 1]> var_9778_cast_fp16 = add(x = var_9776_cast_fp16, y = var_9777_to_fp16)[name = string("op_9778_cast_fp16")];
            fp32 var_9779_epsilon_0 = const()[name = string("op_9779_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 16, 1]> var_9779_cast_fp16 = rsqrt(epsilon = var_9779_epsilon_0, x = var_9778_cast_fp16)[name = string("op_9779_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> var_9780_cast_fp16 = mul(x = x_335_cast_fp16, y = var_9779_cast_fp16)[name = string("op_9780_cast_fp16")];
            tensor<fp16, [1024]> model_layers_27_post_attention_layernorm_weight_to_fp16 = const()[name = string("model_layers_27_post_attention_layernorm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743045952)))];
            tensor<fp16, [1, 16, 1024]> input_275_cast_fp16 = mul(x = var_9780_cast_fp16, y = model_layers_27_post_attention_layernorm_weight_to_fp16)[name = string("input_275_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_gate_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(743048064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746193856))))[name = string("model_layers_27_mlp_gate_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_193_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_gate_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_193_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> var_9785_cast_fp16 = silu(x = linear_193_cast_fp16)[name = string("op_9785_cast_fp16")];
            tensor<fp16, [3072, 1024]> model_layers_27_mlp_up_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [3072, 1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(746200064))), scale = tensor<fp16, [3072, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749345856))))[name = string("model_layers_27_mlp_up_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 3072]> linear_194_cast_fp16 = linear(bias = linear_4_bias_0_to_fp16, weight = model_layers_27_mlp_up_proj_weight_to_fp16_quantized, x = input_275_cast_fp16)[name = string("linear_194_cast_fp16")];
            tensor<fp16, [1, 16, 3072]> input_cast_fp16 = mul(x = var_9785_cast_fp16, y = linear_194_cast_fp16)[name = string("input_cast_fp16")];
            tensor<fp16, [1024, 3072]> model_layers_27_mlp_down_proj_weight_to_fp16_quantized = constexpr_blockwise_shift_scale(data = tensor<int8, [1024, 3072]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(749352064))), scale = tensor<fp16, [1024, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752497856))))[name = string("model_layers_27_mlp_down_proj_weight_to_fp16_quantized")];
            tensor<fp16, [1, 16, 1024]> linear_195_cast_fp16 = linear(bias = linear_1_bias_0_to_fp16, weight = model_layers_27_mlp_down_proj_weight_to_fp16_quantized, x = input_cast_fp16)[name = string("linear_195_cast_fp16")];
            tensor<fp16, [1, 16, 1024]> x_337_cast_fp16 = add(x = x_335_cast_fp16, y = linear_195_cast_fp16)[name = string("x_337_cast_fp16")];
            int32 var_9795 = const()[name = string("op_9795"), val = int32(1)];
            int32 x_batch_dims_0 = const()[name = string("x_batch_dims_0"), val = int32(0)];
            bool x_validate_indices_0 = const()[name = string("x_validate_indices_0"), val = bool(false)];
            string answerRow_to_uint16_dtype_0 = const()[name = string("answerRow_to_uint16_dtype_0"), val = string("uint16")];
            tensor<uint16, [1]> answerRow_to_uint16 = cast(dtype = answerRow_to_uint16_dtype_0, x = answerRow)[name = string("cast_0")];
            tensor<fp16, [1, 1, 1024]> x_cast_fp16_cast_uint16 = gather(axis = var_9795, batch_dims = x_batch_dims_0, indices = answerRow_to_uint16, validate_indices = x_validate_indices_0, x = x_337_cast_fp16)[name = string("x_cast_fp16_cast_uint16")];
            fp16 var_9801_promoted_to_fp16 = const()[name = string("op_9801_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [1, 1, 1024]> var_9803_cast_fp16 = pow(x = x_cast_fp16_cast_uint16, y = var_9801_promoted_to_fp16)[name = string("op_9803_cast_fp16")];
            tensor<int32, [1]> var_9805_axes_0 = const()[name = string("op_9805_axes_0"), val = tensor<int32, [1]>([-1])];
            bool var_9805_keep_dims_0 = const()[name = string("op_9805_keep_dims_0"), val = bool(true)];
            tensor<fp16, [1, 1, 1]> var_9805_cast_fp16 = reduce_mean(axes = var_9805_axes_0, keep_dims = var_9805_keep_dims_0, x = var_9803_cast_fp16)[name = string("op_9805_cast_fp16")];
            fp16 var_9806_to_fp16 = const()[name = string("op_9806_to_fp16"), val = fp16(0x1.1p-20)];
            tensor<fp16, [1, 1, 1]> var_9807_cast_fp16 = add(x = var_9805_cast_fp16, y = var_9806_to_fp16)[name = string("op_9807_cast_fp16")];
            fp32 var_9808_epsilon_0 = const()[name = string("op_9808_epsilon_0"), val = fp32(0x1.197998p-40)];
            tensor<fp16, [1, 1, 1]> var_9808_cast_fp16 = rsqrt(epsilon = var_9808_epsilon_0, x = var_9807_cast_fp16)[name = string("op_9808_cast_fp16")];
            tensor<fp16, [1, 1, 1024]> var_9809_cast_fp16 = mul(x = x_cast_fp16_cast_uint16, y = var_9808_cast_fp16)[name = string("op_9809_cast_fp16")];
            tensor<fp16, [1024]> model_norm_weight_to_fp16 = const()[name = string("model_norm_weight_to_fp16"), val = tensor<fp16, [1024]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752499968)))];
            tensor<fp16, [1, 1, 1024]> last_cast_fp16 = mul(x = var_9809_cast_fp16, y = model_norm_weight_to_fp16)[name = string("last_cast_fp16")];
            tensor<fp16, [151936]> linear_196_bias_0_to_fp16 = const()[name = string("linear_196_bias_0_to_fp16"), val = tensor<fp16, [151936]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(752502080)))];
            tensor<fp16, [1, 1, 151936]> linear_196_cast_fp16 = linear(bias = linear_196_bias_0_to_fp16, weight = model_embed_tokens_weight_to_fp16, x = last_cast_fp16)[name = string("linear_196_cast_fp16")];
            int32 var_9815_axis_0 = const()[name = string("op_9815_axis_0"), val = int32(-1)];
            bool var_9815_keep_dims_0 = const()[name = string("op_9815_keep_dims_0"), val = bool(false)];
            string var_9815_output_dtype_0 = const()[name = string("op_9815_output_dtype_0"), val = string("int32")];
            tensor<int32, [1, 1]> nextToken = reduce_argmax(axis = var_9815_axis_0, keep_dims = var_9815_keep_dims_0, output_dtype = var_9815_output_dtype_0, x = linear_196_cast_fp16)[name = string("op_9815_cast_fp16")];
            int32 var_9822 = const()[name = string("op_9822"), val = int32(0)];
            bool var_9823_interleave_0 = const()[name = string("op_9823_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newKeys = concat(axis = var_9822, interleave = var_9823_interleave_0, values = (k_1_cast_fp16, k_3_cast_fp16, k_5_cast_fp16, k_7_cast_fp16, k_9_cast_fp16, k_11_cast_fp16, k_13_cast_fp16, k_15_cast_fp16, k_17_cast_fp16, k_19_cast_fp16, k_21_cast_fp16, k_23_cast_fp16, k_25_cast_fp16, k_27_cast_fp16, k_29_cast_fp16, k_31_cast_fp16, k_33_cast_fp16, k_35_cast_fp16, k_37_cast_fp16, k_39_cast_fp16, k_41_cast_fp16, k_43_cast_fp16, k_45_cast_fp16, k_47_cast_fp16, k_49_cast_fp16, k_51_cast_fp16, k_53_cast_fp16, k_cast_fp16))[name = string("op_9823_cast_fp16")];
            int32 var_9825 = const()[name = string("op_9825"), val = int32(0)];
            bool var_9826_interleave_0 = const()[name = string("op_9826_interleave_0"), val = bool(false)];
            tensor<fp16, [28, 8, 16, 128]> newValues = concat(axis = var_9825, interleave = var_9826_interleave_0, values = (v_1_cast_fp16, v_3_cast_fp16, v_5_cast_fp16, v_7_cast_fp16, v_9_cast_fp16, v_11_cast_fp16, v_13_cast_fp16, v_15_cast_fp16, v_17_cast_fp16, v_19_cast_fp16, v_21_cast_fp16, v_23_cast_fp16, v_25_cast_fp16, v_27_cast_fp16, v_29_cast_fp16, v_31_cast_fp16, v_33_cast_fp16, v_35_cast_fp16, v_37_cast_fp16, v_39_cast_fp16, v_41_cast_fp16, v_43_cast_fp16, v_45_cast_fp16, v_47_cast_fp16, v_49_cast_fp16, v_51_cast_fp16, v_53_cast_fp16, v_cast_fp16))[name = string("op_9826_cast_fp16")];
        } -> (nextToken, newKeys, newValues);
}